Tolgahan YILMAZ - (IJAEBM) INTERNATIONAL JOURNAL OF ADVANCED ECONOMICS AND BUSINESS MANAGEMENT

Vol No. 1, Issue No. 2, 081 - 092

Improving Portfolio Optimization

by DCC and DECO GARCH:
Evidence from Istanbul Stock Exchange

Tolgahan YILMAZ
Ph.D. Research Scholar
Department of International Finance
Yeditepe University

M
Istanbul, Turkey
[email protected]

Abstract- In this paper, the performance of global minimum

variance (GMV) portfolios constructed by DCC and DECO- Since the mean-variance framework is developed
GARCH are compared to that of GMV portfolios constructed by
by Markowitz [1], there have been profound
EB
sample covariance and constant correlation methods in terms of
reduced volatility. Also, the performance of GMV portfolios are developments on the portfolio optimization. Sharpe
tested against that of equally weighted and cap weighted proposes the CAPM is proposed as the single factor
portfolios. Portfolios are constructed from the stocks listed in model to estimate covariance matrices [2]. Elton and
Istanbul Stock Exchange 30 index (hereafter, ISE-30). The
results show that GMV portfolios constructed by DCC-GARCH Gruber introduce the constant correlation
outperformed the other portfolios. In addition, the performance methodology to reduce the burden of large scale
of GMV portfolios estimated by DCC and DECO-GARCH parameter estimation [3]. Instead of simple model of
methods are improved by extending calibration period from
three years to four years and lowering rolling window term from
Sharpe, multi-factor model estimation is applied by
one week to one day, while the performances of other GMV Chan, Karceski and Lakonishok [4].
portfolios decrease. It shows the effect of time varying variance
and dynamic correlations on portfolio optimization at Turkish
stock market. Standard deviations and the pair-wise
A
Keywords: DCC-GARCH, DECO-GARCH, GMV portfolio, correlations are the elements for covariance
Sample Covariance, Constant Correlation construction. Employing unconditional standard
deviations and constant correlations to estimate
INTRODUCTION
I. covariance matrix are always debated by finance
literature and market as well. The presence of time
Investors and even portfolio managers often varying variances and correlations are shown by
monitors the benchmark indices which are generally
IJ

Engle [5]. GARCH (1,1), introduced by Bollerslev,

constructed by exchanges relying on cap weighted capture the ARCH effect and model the time
average. These benchmarks only give the investors varying variance with less constraints relative to
the general idea about the general market ARCH [6].
movement. Since they lack the requirements for
robust benchmark or portfolio construction,
investments decisions depend on these benchmarks The first adaptation of univariate GARCH
lead the investors and portfolio managers to process is carried out by Bollerslev, Engle and
underperform. Therefore, robust portfolio Wooldridge [7]. They employ the univariate
optimization is one of the major issues for portfolio GARCH process to do multivariate
managers and other market participants. parameterization. It is known as vech form of

1
ISSN: 2230-7826 @ 2011 http://www.ijaebm.iserp.org. All rights Reserved. Page 81
 Tolgahan YILMAZ - (IJAEBM) INTERNATIONAL JOURNAL OF ADVANCED ECONOMICS AND BUSINESS MANAGEMENT
Vol No. 1, Issue No. 2, 081 - 092

multivariate GARCH. They also propose diagonal Tse and Tsui support estimation accuracy of
vech form of that model by which the numbers of dynamic correlation model. They assume that the
parameters, which are to be estimated, were reduced. pair-wise correlations follow moving average
However, it creates a computational burden when process and they find the conditional variances by
the sample size increases. Since the number of univariate GARCH [13]. They show that errors of
parameters to be estimated is too many, it is hard to maximum likelihood estimator are reduced by the
achieve a feasible estimation. multivariate GARCH estimation with dynamic
correlation.

Engle and Kroner find the way of producing

positive definite covariance matrix through BEKK Considering that DCC-GARCH captures the time
model [8]. However, it is a great problem that varying correlations and variances, it is very well
estimating the conditional covariances as the sample structured model to estimate time varying
size increased. In order to solve this problem, covariance matrix. However, the estimation of
Bollerslev introduces constant conditional conditional correlation matrix for a portfolio with

M
correlation GARCH (CCC-GARCH) model [9]. By large number of assets causes the difficulty of
this model, standard deviations of each asset are estimation. The way of reducing the scale of
produced by univariate GARCH process and it is estimation was proposed by Engle and Kelly.
assumed that the pair-wise correlations are constant. Averaging of pair dynamic correlations, they reduce
Therefore, CCC approach does not model the time the burden of large scale parameterization. This
varying conditional correlation. The standard process is called as Dynamic Equicorrelation
deviations within the covariance matrix are GARCH (DECO-GARCH)[14]. It reduces the
calculated relying on the GARCH constraints such sample risk caused by large scale covariance matrix.
EB
as non-negativity. So, under the conditions of the However, there is always estimation risk because of
guaranteed non-negative conditional variances and assigning one value to each pair-wise correlation
the invertible conditional matrix generating positive instead of their real values.
definite covariance matrix is certainly obtained.

These relatively new techniques, DCC and

However, Tse and Yu prove that the constant DECO-GARCH, are employed to construct GMV
correlation is not valid when the estimation process portfolio in this paper. Considering the high volatile
is multivariate [10]. The pair-wise correlations are structure of emerging stock markets, one of the
also time-varying and they need to be modeled to volatile emerging market indexes, Istanbul Stock
produce consistent errors. Exchange 30 Index (ISE-30) and its constituents are
A
used to test the performance of GMV.

The challenging problem of constant correlation

is solved by the dynamic conditional correlation In order to test the effect of time varying variance
GARCH (DCC-GARCH), proposed by Engle [11]. and dynamic correlations on portfolio optimization,
Mathematical framework of this model has main the estimation accuracy of these methods is
IJ

two steps algorithm to have time varying covariance compared to that of the sample covariance and
matrix. First step is to find conditional standard constant correlation methodology in terms of
deviations through the univariate GARCH and reduced volatility of the GMV portfolios. The
second step is to model the time varying correlations performances of non-optimized portfolios such as
relying on lagged values of residuals and covariance equally weighted and cap weighted portfolios are
matrices. After that, conditional covariance matrix also compared to that of those GMV portfolios.
could be found by using conditional standard
deviations and dynamic correlations [12].
The rest of the paper is organized as following:
Section 2 gives brief information about theoretical
framework of mean-variance optimization and

2
ISSN: 2230-7826 @ 2011 http://www.ijaebm.iserp.org. All rights Reserved. Page 82
 Tolgahan YILMAZ - (IJAEBM) INTERNATIONAL JOURNAL OF ADVANCED ECONOMICS AND BUSINESS MANAGEMENT
Vol No. 1, Issue No. 2, 081 - 092

covariance estimation techniques employed in this Lagrange multiplier (L) with respect to weight is
paper. While Section 3 describes the data and equaled to zero such that
evaluates the empirical results, concluding remarks
are finally given at Section 4
L
1
2

ww  1 we  1  2 wr  r p 
L
 w  1e  2 r  0
II. METHODOLOGY w

A. Estimating Covariance Matrix by Constant

Correlations and Volatilities w  1 1e  2  1r 

As modern portfolio theory (MPT) proposes, Now, it is also convenient to write portfolio
main objective of diversification is to minimize risk variance in terms of multipliers ( 1 and  2 ) such
in a given level of return. While all efficient that
portfolios nest on the efficient frontier, GMV is the

M
one at the beginning of the efficient frontier and it
has lowest volatility amongst other efficient  2p  ww

portfolios.  
 2p  w 1 1e  2 1r  1w1e  2 w1r
 2p  1we  2 wr  1  2rp
Considering mean variance optimization, object
function is portfolio variance and constraints are
following; weights of assets in the portfolio must be  2p  1  2 rp  1  2 wr 
add up to one and the combination of asset returns
EB
must be equal to target portfolio return. In-mean So, the equation of portfolio variance enables us
variance framework, efficient frontier is constructed to find the level of variance given expected portfolio
by portfolios that gives minimum variance amongst return, if the multipliers are known. It means that
other portfolios in the feasible region at the given equation above is also a tool for constructing
expected portfolio return. Constructing efficient efficient set.
frontier requires quadratic optimization program to
find optimal weights, which construct efficient
portfolio, is given by Since we have two unknowns in the weight
equation and so in portfolio variance, we can use
two constraints to find 1 and  2 . We multiply
1
A
min ww subject to weight equation with transpose of vector of ones and
w 2
w r  r p  vector of expected asset returns. Since, values of
constraints are known, now it is trivial to solve
w e  1
system of equation below.
where
IJ

ew  1e 1e  2 e 1r  1


1
 w1 
 11   1n 
 r1 
1 r w  1r  1e  2 r  1r  r p
w  r   
  2  en1   
 
wn1   2   nn       rn1
       It is possible to write the system above as simple
   n1   n1    
 wn  rn  1 system of linear equations without mentioning
matrix notation such that [15]
The Lagrange Multiplier method can be used to
find optimum weights that minimize the portfolio
1  2  1 
variance. First, the constraints are added to the 1   2  r p
object function and then the first derivative of the

3
ISSN: 2230-7826 @ 2011 http://www.ijaebm.iserp.org. All rights Reserved. Page 83
 Tolgahan YILMAZ - (IJAEBM) INTERNATIONAL JOURNAL OF ADVANCED ECONOMICS AND BUSINESS MANAGEMENT
Vol No. 1, Issue No. 2, 081 - 092

where Then, we can write the terms in the equation

above in the form of vector and matrices.
  e 1e ;   e 1r ;   r  1r 
Since wr  rp
1 1
Since e r  r  e , we denote both of them as  .
After solving that simple system for 1 and  2 , we e  1r e  1r e  1
wr   w   r 1  w  
have the results such that e  1e e  1e e  1e

Finally, vector of optimal portfolio weights for

   rp ;
the GMV portfolio are presented by
 rp  
1  2 2 
      2
 1e
Now we substitute those values into the equation wgmv  
of portfolio variance such that e 1e

Aftermath, an important question may arise such

M
 2p 
   rp

 rp  
rp 
that do we really need to have constraint related with
   2    2 expected portfolio return to find optimum weights of
GMV portfolio? Since there is no term related to
expected asset returns or expected portfolio return,
2  2 r
  2p     r p p  the intuitive answer for that question could be “No”.
   2 To prove that we use the same object function that is
portfolio variance and only one constraint that is
sum of all weights must add up to 1. New nonlinear
Actually, we have equation for optimal weights
EB
for minimum variance portfolios given expected optimization program to minimize portfolio variance
asset returns and portfolio return on the efficient set can be stated such that;
such that
1
min w  w
w
r  1r  rp e 1r
 1e 
rp e 1e  e 1r
 1r
 w 2
e er  r  e r 
1 1 1 2
e er  r  e r 
1 1 1 2 s.t. w e  1
1
However, in this study we search for minimum L w w   w e  1 
2
point of the efficient frontier. Put it differently, we L
need to find the optimum weights of a portfolio  w   e  0
A
w
which gives the minimum variance compared to
other efficient portfolios on the efficient set and w   1e
therefore, it is called global minimum variance We can substitute that value of weight into the
(GMV) portfolio. The critical calculus takes a place constraint such that
here in order to find optimal weights for GMV
portfolio explicitly. First step is taking derivation of
IJ

portfolio variance (  2p ) with respect to expected


 
we  1    1e e  1 

portfolio return ( r p ) and then equating the 

  e e1 1

derivation to zero. First order derivation of portfolio Now, we can replace the value of  into the
variance with respect to expected portfolio return is weight equation. So, it gives us the equation of
given below; optimum weights for GMV portfolio such that

d 2p 2 r p  2  
  0  rp 
dr p    2 

4
ISSN: 2230-7826 @ 2011 http://www.ijaebm.iserp.org. All rights Reserved. Page 84
 Tolgahan YILMAZ - (IJAEBM) INTERNATIONAL JOURNAL OF ADVANCED ECONOMICS AND BUSINESS MANAGEMENT
Vol No. 1, Issue No. 2, 081 - 092


w  e 1e 11e matrix and the matrix constructed by multiplication
of vectors of constant standard deviations such that

 1e  cc  S S  R . This technique is named as constant
wgmv 
e 1e correlation estimation and developed by Elton and
Gruber (1973). Aftermath, finding optimal weights
Finally, it is proved that there is no need to have a for GMV portfolio is the same with sample
constraint related to expected portfolio return, if we covariance such that
try to find optimum weights for GMV portfolio.
Relying on that result, the quadratic program above
can be solved analytically. 1
 cc 1
wgmv  1

1 cc 1
So far, there are no specific constraints on
portfolio weights. If a constraint is imposed on
B. Dynamic Conditional Correlation Model (DCC-
portfolio weights such that wi0 , the optimization
GARCH)
problem can not be solved analytically anymore. If

M
there are inequality constraints in the optimization
problem, it requires the numerical optimization. Being a multivariate GARCH model, DCC-
Since the short sale constraint is the case for that GARCH assumes that returns of the assets ( rt )
paper, Matlab “frontcon” function is employed to distribute normally with zero mean and they have
solve minimization portfolio variance problem covariance matrix such as Ct.
numerically given inequality constraints of weights.
rt ~ N 0, Ct  
EB
The key part is in equation of w gmv is finding the
Conditional covariance matrix is found by using
covariance matrix denoted as  . When constant
conditional standard deviations and dynamic
standard deviations and pair-wise correlations are
correlation matrix. Let St is 1 n vector of
used to estimate covariance matrix, this optimization
conditional standard deviations modeled by
procedure is called sample covariance estimation.
univariate GARCH process such that
Considering sample covariance estimation,
  S S  R where R is n by n square matrix of
constant correlations and S refers to the vector of
  j t2 j 
p q
S t2   0    i et2i 
constant standard deviations such as S  s1 , s 2,..., sn   i 1 j 1

where s i =Standard deviation of ith asset in the

where et i   t i  t i and  ti ~ N 0,1 . In order to
A
portfolio.
find time varying correlation matrix, Engle
proposes a model [11] for the time varying
In constant correlation estimation, R is covariance such that
constructed by using average of pair-wise
correlations such that p q p
 
K t  (1    i    j ) K    i et i eti    j K t  j 
q
IJ

i 1 j 1 i 1 j 1

 1   
     
R    1nn  1     I nn R  s.t  i  0
    
  j 0
    1
p q

where  is average constant correlation, 1nn is i    j  1

i 1 j 1
n  n matrix of ones and I nn is n  n identity matrix.
As it can be seen from that equation above, the
Then, new covariance matrix is estimated by GARCH process is adopted to model time varying
Hadamard product of constant average correlation

5
ISSN: 2230-7826 @ 2011 http://www.ijaebm.iserp.org. All rights Reserved. Page 85
 Tolgahan YILMAZ - (IJAEBM) INTERNATIONAL JOURNAL OF ADVANCED ECONOMICS AND BUSINESS MANAGEMENT
Vol No. 1, Issue No. 2, 081 - 092

covariance matrices. K is the unconditional S i , j ,t

 i, j ,t 
covariance and it is initially obtained by sample S ii2 S 2jj
covariance estimation. K t is forecasted by lagged 
residuals ( et  i ), which are standardized by Finally, conditional covariance matrix can be
conditional standard deviations, and lagged found by Hadamard product of the matrix of
covariances ( K t  i ). Therefore, in estimating conditional standard deviations and time varying
conditional covariance matrix first, conditional correlation matrix such that
standard deviations of each asset in the portfolio are
modeled by univariate GARCH. It is important to
Ct  S t  S t  M t 
note that the constraints of GARCH process are still
considered to construct positive definite covariance
matrix. In order to find estimators of that model, the This methodology gives the conditional
log likelihood function can be written as covariance matrix for each data point of the
calibration period. To find the conditional
covariance matrix that is used to optimize weights of

M
L
1 T

 k log 2   log  S t M t S t   rtS t1M t1S t1rt  assets in the portfolio, one day conditional standard
2 t 1 deviations of assets and their dynamic correlation

1 T

 k log 2   2 log  S t   log  Rt   et M t1et
2 t 1
 matrix are forecasted.

C. Dynamic Equicorrelation Model (DECO-

After finding optimum estimators that maximize GARCH)
the log likelihood function above, it is easy to
produce covariance series. But, it is necessary to
Engle and Kelly propose a different version of
note that each covariance matrix is not constructed
EB DCC-GARCH model, named DECO-GARCH [14].
by conditional standard deviations yet. This
They set the average of conditional correlation equal
covariance matrix series is generated by relying on
to all pair correlations in order to reduce burden of
initial unconditional covariance matrix. Then new
the computation of large scale correlation matrices.
covariance matrix for next time point is generated by
They use the same structure to construct covariance
previous one and standardized residuals as in simple
matrix as in the DCC-GARCH model such that
univariate GARCH process. So, univariate GARCH
process is employed to extract time varying positive
definite correlation matrices from that covariance Ct  S t S t  M t 
matrix series such that
However, the conditional correlation matrix
A
1 1
 would be different because of taking average of
M t   K td  K t  K td 
    conditional correlations as the below equation shows

In that equation, K td refers to the diagonal matrix 1 n 

of variances which are obtained from the diagonal
t   ˆ ij,t
nn  1 i , j 1
items of K t as following i j
IJ

2
 S11 0  0  where  t is defined as conditional equicorrelation as
 
0 2
 0  ̂ ij,t refers to the pair-wise correlation. After finding

S 22
K td
   

 
2
 the average correlation, the new conditional
 0 0  S nn  correlation matrix is constructed such that
The matrix notation for time varying correlation
indicates that the way of calculating correlations M t   t 1nn  1   t I nn 
such as dividing covariances by standard deviations
extracted from the covariance matrix. The matrix That equation is almost same with the equation
notation can be interpreted by algebraically such that that calculates the constant correlation matrix. The

6
ISSN: 2230-7826 @ 2011 http://www.ijaebm.iserp.org. All rights Reserved. Page 86
 Tolgahan YILMAZ - (IJAEBM) INTERNATIONAL JOURNAL OF ADVANCED ECONOMICS AND BUSINESS MANAGEMENT
Vol No. 1, Issue No. 2, 081 - 092

only difference is that correlation matrix is extracted covariance matrix with a calibration period of three
from covariance matrix series modeled by univariate years is performed by ruling out the short sale and
GARCH process and therefore, average correlation then, 1 week out-of sample forecasting of optimal
matrix in that equation is conditional and time portfolio returns are generated. Calibration period of
varying. three years is extended to four years to test the effect
of changing calibration period on the performance of
the GMV portfolio. Then, in order to make the
The log- likelihood function is defined by [14] portfolio allocation more dynamic, the out-of sample
forecasting period is lowered to 1 day.

 t t
log 1   n 1 1  n  1  

1   2   Since the list of constituents of the index changes
L   
1 
 
2 t
 1     ei, t  1  n  1
T t 
t  i

 e  
  i , t   over time, a fixed list of 23 stocks within the ISE-30
 t  i  
   Index dating from January 2004 is used to simulate a
cap-weighted portfolio for fair comparison. Since
Considering DCC, it is necessary to estimate
there are some short sale restrictions at Istanbul

M
n  n  1
pair-wise correlations. Instead of estimating
2 Stock Exchange, the short sale constraint is added to
each pair-wise correlation, only one parameter for constraint set of optimizer.
conditional correlation is estimated by DECO. When
the sample size is quite large, the DCC model incurs B. Emprical Results
sample size risk due to the fact that the number of
parameters which are to be estimated would be Reduced volatility is the major criteria to test the
many. It creates a burden on programming and also performances of GMV portfolios. First, the out of
noise on the forecasted data. Considering DECO sample returns within 1 week rolling window period
approach, the computational burden may be
EB is forecasted by using calibration period of three
decreased by assigning same correlation to the each years. Then, the forecasting period is lowered from
element of conditional correlation matrix but, there one week to one day. So, Out-of sample returns are
would be model risk for the sample in which each forecasted by using 3 year calibration period from
asset returns has quite different dynamic pair-wise the beginning of January 2008 till the end of
correlations. So, this paper also answers that September 2010. The Table 1 and 2 show the risk
question whether the pair-wise correlations of and return measures of the GMV portfolios,
Turkish stocks are not quite different from each generated by different covariance estimation
other to be negligible or not. If the results indicate methods with three year calibration period, for one
that performance of GMV portfolio estimated by week and one day rolling windows. Also the
DECO is better than that of GMV portfolio performance of the each GMV is compared to that
A
constructed by DCC, it can be concluded that pair- of equally weighted and cap weighted portfolios. On
wise correlations move very close to mean. the tables, SC, EQW, CAPW, CC refer to sample
Otherwise, this is not the case. covariance estimation, equally weighted portfolio,
cap-weighted portfolio and constant correlation
covariance estimation respectively.
IJ

III. DATA AND EMPRICAL RESULTS

A. Data

In the paper, the data, composed of the daily

prices of stocks within the ISE-30 Index from
January 2004 to the end of the September 2010, is
collected from FOREX FX2000 6.1.233. Then, daily
prices are transformed to daily logarithmic returns.
A rolling-window estimate of the variance-

7
ISSN: 2230-7826 @ 2011 http://www.ijaebm.iserp.org. All rights Reserved. Page 87
 Tolgahan YILMAZ - (IJAEBM) INTERNATIONAL JOURNAL OF ADVANCED ECONOMICS AND BUSINESS MANAGEMENT
Vol No. 1, Issue No. 2, 081 - 092

TABLE I. GMV PORTFOLIO-RISK AND RETURN STATISTICS and other portfolios for four year calibration period
Covariance Estimation Methods with one week and one day rolling windows.
Statistics (Calibration:3 Years Roll Window: 1Week)
SC EQW CAPW CC DCC DECO
Mean TABLE III. GMV PORTFOLIO-RISK AND RETURN STATISTICS
0.058% 0.055% 0.073% 0.061% 0.031% 0.03%
Return
Covariance Estimation Methods
Annualized
Statistics (Calibration:4 Years Roll Window: 1Week)
Mean 15.65% 15.00% 20.29% 16.70% 8.06% 7.99%
SC EQW CAPW CC DCC DECO
Return
Mean - -
Volatility 1.399% 1.493% 1.659% 1.598% 1.411% 1.629% 0.009% 0.020% 0.040% 0.037%
Return 0.035% 0.011%
Annualized Annualized
Volatility 22.22% 23.70% 26.34% 25.37% 22.39% 25.86% - -
Mean 2.383% 5.227% 10.73% 9.892%
8.508% 2.831%
Skewness Return
-0.320 -0.227 -0.083 -0.208 -0.399 -0.489
Excess Volatility 1.490% 1.584% 1.725% 1.668% 1.491% 1.723%
Kurtosis 1.788 1.681 1.426 2.870 5.041 6.004 Annualized
23.65% 25.15% 27.38% 26.49% 23.67% 27.36%
Historical Volatility
VaR (99%) 4.760% 4.863% 5.226% 5.476% 4.670% 6.413%
Skewness -0.258 -0.176 -0.119 -0.196 -0.481 -0.706

TABLE II. GMV PORTFOLIO-RISK AND RETURN STATISTICS Excess

1.410 1.442 1.425 2.317 4.382 5.865

M
Kurtosis
Covariance Estimation Methods Historical
4.739% 4.872% 5.987% 5.616% 5.298% 7.256%
Statistics (Calibration:3 Years Roll Window: 1Day) VaR (99%)
SC EQW CAPW CC DCC DECO
Mean TABLE IV. GMV PORTFOLIO-RISK AND RETURN STATISTICS
0.060% 0.057% 0.075% 0.063% 0.050% 0.046%
Return
Covariance Estimation Methods
Annualized
Statistics (Calibration:4 Years Roll Window: 1Day)
Mean 16.21% 15.54% 20.92% 17.27% 13.40% 12.32%
Return SC EQW CAPW CC DCC DECO
Mean
Volatility 1.393% 1.491% 1.657% 1.595% 1.374% 1.606% 0.011% 0.021% 0.041% 0.068% 0.031% 0.066%
Return
Annualized
Volatility
22.12% 23.67% 26.31%
EB
25.32% 21.82% 25.50% Annualized
Mean 2.732% 5.506% 10.94% 18.61% 8.01% 18.04%
Skewness -0.323 -0.230 -0.086 -0.226 -0.346 -0.519 Return
Excess Volatility 1.487% 1.583% 1.724% 1.605% 1.153% 1.340%
1.747 1.691 1.433 2.795 4.712 5.715
Kurtosis Annualized
23.61% 25.13% 27.36% 25.47% 18.30% 21.28%
Historical Volatility
4.672% 4.863% 5.226% 5.488% 4.745% 6.303%
VaR (99%)
Skewness -0.253 -0.178 -0.120 -0.620 -1.062 -1.617
Excess
According to the results, volatility of the Kurtosis
1.414 1.446 1.431 3.209 9.657 14.559
portfolios decreases as the forecasting period Historical
4.695% 4.872% 5.987% 5.948% 4.131% 4.995%
VaR (99%)
(rolling window) is lowered. GMV portfolio
constructed by DCC-GARCH produce the lowest According to the tables above, volatility of all
volatility while portfolio optimization is carried out
A
portfolios except ones optimized by DCC and
by more dynamically (when rolling window is one DECO-GARCH is raised by extending the
day). calibration period from three years to four years.
The performances of the portfolios estimated by
DCC and DECO-GARCH increases substantially in
In addition, all portfolio returns are negative
terms of reduced volatility when four year
skewed and the distribution of the portfolio returns
IJ

calibration period accompanied with one day rolling

shows fat-tail since all portfolios have positive
window is employed. DCC-GARCH covariance
excess kurtosis. Value at Risk Analysis is also
estimation gives lowest volatility amongst all
employed. DCC-GARCH and sample covariance
portfolios.
give lowest historical VaR for three year calibration
period.
Historical VaR analysis also supports this result,
since the lowest historical VaR is achieved by DCC
To test the effect of using larger scale of data, the
process having four year calibration and one day
calibration period is extended. Table 3 and 4
rolling window periods.
indicate the risk and return measures of the GMV

8
ISSN: 2230-7826 @ 2011 http://www.ijaebm.iserp.org. All rights Reserved. Page 88
 Tolgahan YILMAZ - (IJAEBM) INTERNATIONAL JOURNAL OF ADVANCED ECONOMICS AND BUSINESS MANAGEMENT
Vol No. 1, Issue No. 2, 081 - 092

Visualizing changes on conditional volatilities The conditional volatility is modeled by univariate

over forecasting time horizon is another way to EGARCH(1, 1).
compare the performances of GMV portfolios in
terms of volatility. Figure 1 and 2 present the
conditional volatilities of each portfolio for each
calibration period and the period of rolling window.

Figure 1. Conditional Volatility-Calibration perıod:3 Years

M
EB
Figure 2. Conditional Volatility-Calibration period:4 Years
A
IJ

Considering one day rolling window period, the DCC with four year calibration and one day rolling
conditional volatility series of portfolios window.
constructed by DCC is below the volatilities of
other portfolios for each calibration period.
Although conditional volatility of sample In order to analyze the estimation quality of
covariance moves at lower level than DCC as one methods, we compared the performance of forecasted
week rolling window period is employed, the GMV portfolios to the performance of the true GMV
lowest conditional volatility level is achieved by portfolios (in-sample estimation). Since the forecasted
GMV portfolio with lowest volatility is reached by
out-of sample estimation with four year of calibration

9
ISSN: 2230-7826 @ 2011 http://www.ijaebm.iserp.org. All rights Reserved. Page 89
 Tolgahan YILMAZ - (IJAEBM) INTERNATIONAL JOURNAL OF ADVANCED ECONOMICS AND BUSINESS MANAGEMENT
Vol No. 1, Issue No. 2, 081 - 092

period, in-sample estimation for GMV portfolios is DCC-GARCH and DECO-GARCH. Those efficient
carried out by using realized returns of forecasting frontiers are given by figure below:
period starts from January 2008 to September 2009.
Four different efficient frontiers are drawn by using
four different covariance estimation methods,
which are sample covariance, constant correlation,

Figure 3. In -Sample Estimation-Mean Variance Efficient Frontier

M
EB
Relying on the in-sample estimation, lowest According to Table 5, the portfolio that is closest to
volatility is achieved by DCC-GARCH and true GMV portfolio is the one constructed by DCC-
DECO-GARCH is the second one. Forecasting GARCH estimation. While DECO estimation follows
A
accuracy can be measured by spread between the DCC, other estimation methods create substantial
volatility of forecasted GMV and that of true deviations from true volatilities. The spreads generated
GMV. To quantify this graphical analysis, by dynamic correlations models are almost 25%-40%
volatilities of forecasted GMV and true GMV lower than the spreads produced by other estimation
portfolios are given table below: methods. Therefore, it can be stated that GMV
portfolios with dynamic correlations outperform the
IJ

TABLE V. IN-SAMPLE ESTIMATION other portfolios in terms of reduced risk.

Covariance Forecasted True
Spreada
Estimation Methods GMV Vol. GMV Vol.
Sample Covariance 1.49% 0.89% 0.60%

Constant Correlation 1.60% 0.97% 0.63%

DCC 1.15% 0.77% 0.38%
DECO 1.34% 0.88% 0.46%
a. Spread=|Forecasted GMV Volatility-True GMV Volatility|

10
ISSN: 2230-7826 @ 2011 http://www.ijaebm.iserp.org. All rights Reserved. Page 90
 Tolgahan YILMAZ - (IJAEBM) INTERNATIONAL JOURNAL OF ADVANCED ECONOMICS AND BUSINESS MANAGEMENT
Vol No. 1, Issue No. 2, 081 - 092

Relying on the results of in-sample estimation

IV. CONCLUSION for four year period, the GMV portfolios
constructed by DCC and DECO-GARCH are closer
Employing GARCH process in mean-variance to the true GMV than the other portfolios. The
optimization framework has been the subject of the spread between volatility of forecasted GMV
finance literature for a long time. Considering the portfolio and that of true GMV portfolio is detected
co-movements of assets in a portfolio, the and it is found that lowest spread belongs to DCC.
multivariate form of GARCH process is employed to DECO is the second one. Moreover, volatility
have efficient covariance estimators. spreads of DCC and DECO are almost 25% and
40% lower than those of other covariance
estimation methods. In-sample estimation results
In this paper, the estimation accuracy of two support out-of sample estimation findings related to
different types of multivariate GARCH models, better performance of DCC relative to that of other
DCC and DECO-GARCH, are compared to classical estimation methods under the conditions of more

M
sample covariance and constant correlation dynamic allocation and extended calibration period.
estimation. Also, performances of the forecasted
GMV portfolios are compared to equally weighted
portfolio and cap-weighted portfolio in terms of Finally, there is a necessary question is why
reduced volatility. Volatility series of each portfolio DECO performs worse than DCC and sample
is extracted by univariate EGARCH process due to covariance, except in the case of four year
the fact that the presence of time varying variance. calibration period with one day rolling window?
Also, estimation accuracy of each GMV portfolio is Answer for that question is actually one of the
EB
tested by employing in-sample estimation. major results related to the structure of Turkish
stock market. Before answer that question, digging
up empirical results would provide logical path to
According to the results of out of sample forecast, answer that question precisely. If the performance
when one-day portfolio allocation is used, extending of GMV optimized by constant correlation method
calibration period from three years to four years is detected, it is easy to find it is the worst GMV
reduce the volatility of GMV portfolios constructed portfolio for both out-of and in-sample estimation.
by DCC and DECO. However, this is not case for
other covariance estimation methods such as sample
covariance and constant correlation portfolios. Bottom line, DECO and constant correlation
methods perform worse than the others especially
A
when three year calibration period is used. This
It is important to note that performance of GMV means the estimation error incurred by these two
optimized by DECO is still poor for three years methods are higher than that of the other methods.
calibration period, even if its performance is Reason causing high estimation error is variance of
improved by lowering rolling window term from one pair-wise correlations. Therefore, it can be
week to one day. Also, it performs quite poor with concluded that pair-wise correlations among
IJ

four year calibration and weekly roll. DECO is Turkish stocks do not cluster closely around mean,
second best only under the conditions of four year when the number of data point is less than 1000.
calibration and daily roll. However, DCC estimation Considering three year calibration period,
is the best one in terms of reduced volatility when differences among those correlations are not
allocation period is reduced from one week to one negligible.
day for each calibration term. Also, the lowest
volatility is achieved by DCC when calibration
period is four years and rolling window term is one However, when longer calibration term such as
day. four years and more dynamic allocation period such
as one day is considered, DECO is the second best

11
ISSN: 2230-7826 @ 2011 http://www.ijaebm.iserp.org. All rights Reserved. Page 91
 Tolgahan YILMAZ - (IJAEBM) INTERNATIONAL JOURNAL OF ADVANCED ECONOMICS AND BUSINESS MANAGEMENT
Vol No. 1, Issue No. 2, 081 - 092

GMV. This is the clue for another important [13] Y.K. Tse and A.K.C. Tsui “A Multivariate GARCH Model with Time-
Varying Correlations”, Journal of Business and Economic Statistics,
inference for Turkish stock market. Pair-wise vol. 20, pp. 351–362, 2002.
correlations also change over time, when extended [14] R.F. Engle and B.T. Kelly, “Dynamic Equicorrelation”, NYU Working
calibration period is used such as four year period. If Paper No. FIN-08-038, 2009.
[15] F. Fabozzi, S. Focardi and P.Kolm, “Quantitative Equity Investing:
more dynamic portfolio allocation is preferred, Techniques and Strategies”, John Wiley and Science Inc., 2010,
techniques that model dynamic correlations such as pp.318-319.
DCC and DECO-GARCH should be used.

Covariance estimation techniques that employ

each pair-wise correlations, incur sampling error
because of large number of estimated parameters.
However, the benchmark index, which is
investigated in this paper, consists of only twenty
three stocks for the investigation period. Therefore,

M
the scale of covariance matrices estimated by each
method are not too large and so do not create too
much noise causing sampling error. Consequently,
DCC outperforms other estimation techniques in
terms of reduced volatility, when longer calibration
term and shorter rolling window period is
considered.
EB
REFERENCES
[1] H. Markowitz, “Portfolio Selection”, Journal of Finance, vol.7, pp.77–
91, 1952.
[2] W. Sharpe, “A Simplified Model for Portfolio Analysis. Management
Science”, vol.9, pp.277–93, 1963.
[3] E. Elton and M. Gruber, “Estimating the Dependence Structure of Share
Prices–Implications for Portfolio Selection”, Journal of Finance, vol.28,
pp. 1203–32, 1973.
[4] L. Chan, J. Karceski and J. Lakonishok, “On Portfolio Optimization:
Forecasting Covariances and Choosing the Risk Model”, Review of
Financial Studies, vol.12, pp. 937–74, 1999.
[5] R.F. Engle, “Autoregressive Conditional Heteroscedasticity with
A
Estimates of the Variance of United Kingdom Inflation”, Econometrica,
vol. 50, pp. 987-1007, 1982.
[6] T.Bollerslev, “Generalized Autoregressive Conditional
Heteroskedasticity”, Journal of Econometrics, vol. 31, pp. 307-327,
1986.
[7] T.Bollerslev, R.F.Engle and J.M.Wooldridge, “Capital Asset Pricing
Model with Time-Varying Covariances”, Journal of Political Economy,
vol. 96, pp. 116–131, 1988.
IJ

[8] R.F. Engle and F.K. Kroner, “Multivariate Simultaneous Generalized

ARCH” Econometric Theory, vol. 11, pp. 122-150, 1995.
[9] T.Bollerslev, ”Modeling the Coherence in Short-run Nominal Exchange
Rates: A Multivariate Generalized ARCH Model”, Review of Economics
and Statistics, vol.72, pp. 498–505, 1990.
[10] A.K.Tsui, and Q. Yu, “Constant Conditional Correlation in a Bivariate
GARCH model: Evidence from the stock market in China. Mathematics
and Computers in Simulation”, vol.48, pp. 503–509, 1999.
[11] R.F. Engle, “Dynamic conditional correlation - A simple class of
multivariate GARCH models”, Journal of Business and Economic
Statistics, vol.20, pp. 339-350, 2001.
[12] R.F. Engle and K. Sheppard, “Theoretical and Empirical properties of
Dynamic Conditional Correlation Multivariate GARCH”, NBER
Working Papers 8554, 2001.

12
ISSN: 2230-7826 @ 2011 http://www.ijaebm.iserp.org. All rights Reserved. Page 92