The Role of Covariance Matrix Forecasting Method in the Performance of Minimum-Variance Portfolios

Valeriy Zakamulin This revision: March 19, 2014

Abstract Providing a more accurate covariance matrix forecast can improve substantially the performance of optimized portfolios. In this paper we evaluate dierent covariance matrix forecasting methods by looking at their ability, in out-of-sample tests, to optimize risk and improve risk-adjusted performance of minimum-variance portfolios. We nd large dierences between the methods. Our results suggest that shrinkage of sample covariance matrix improves neither the forecast accuracy nor the performance of minimum-variance portfolios. We demonstrate that the simple exponentially weighted covariance matrix forecast produces superior performance compared to the commonly used sample covariance matrix. Our ndings also reveal that the exponentially weighted covariance matrix forecast usually produces the best risk-adjusted performance of minimum-variance portfolios. Yet, when it comes to the portfolio risk optimization, multivariate GARCH models generally demonstrate the best results. Specically, switching from the sample covariance matrix forecast to a multivariate GARCH forecast reduces the forecasting error and portfolio tracking error by half at least. Key words: covariance matrix forecasting, minimum-variance portfolio optimization, sample covariance, covariance shrinkage, exponentially weighted covariance, multivariate GARCH, model comparison JEL classication: C30, C52, G11, G17

The author is grateful to Steen Koekebakker for his insightful comments. All other comments are welcomed. a.k.a. Valeri Zakamouline, School of Business and Law, University of Agder, Service Box 422, 4604 Kristiansand, Norway, Tel.: (+47) 38 14 10 39, E-mail: Valeri.Zakamouline@uia.no

Electronic copy available at: http://ssrn.com/abstract=2411493

Introduction

There is an ongoing discussion on whether the optimized portfolios can outperform equally weighted portfolios. Whereas DeMiguel, Garlappi, and Uppal (2009) and Duchin and Levy (2009) show that optimized portfolios cannot consistently beat equally weighted portfolios, Kritzman, Page, and Turkington (2010) demonstrate the superiority of optimized portfolios. Specically, the latter researchers nd that portfolios that optimize the mean-variance tradeo show better performance than equally weighted portfolios. What is especially noteworthy is that their ndings reveal that portfolios that minimize the variance of returns outperform the portfolios that optimize the mean-variance tradeo. Clarke, de Silva, and Thorley (2006) were the rst to document the superiority of minimumvariance portfolios. The advantage of minimum-variance portfolios over mean-variance optimized portfolios is two-fold. First, minimum-variance portfolios deliver superior risk-adjusted performance compared to that of mean-variance optimized portfolios. Second, while the implementation of the mean-variance model requires forecasting both the mean returns and covariance matrix of returns, the implementation of the minimum-variance model requires forecasting the covariance matrix of returns only. There is a common consensus that mean returns are notoriously dicult to forecast, whereas forecasting the covariance matrix can be easily done using a rolling sample covariance matrix. A standard approach is to use the monthly rolling n-year covariance matrix, where n varies from 5 to 20 years, as a forward-looking estimate of future covariance matrix (see Chan, Karceski, and Lakonishok (1999), DeMiguel et al. (2009), Duchin and Levy (2009), and Kritzman et al. (2010) among others). Yet, this is a valid approach if one assumes that the covariance matrix is either constant over time or varying very slowly over time. This assumption does not seem to hold for nancial asset returns. On the contrary, there is a large strand of literature that demonstrate that nancial asset returns exhibit heteroskedasticity with volatility clustering. The assumption of constant correlation between nancial asset returns also seems to be violated. As a motivation, Figure 1 plots the monthly realized standard deviations of returns on the US stock market and bond market indices as well as the monthly realized correlation coecient between the returns on these indices. The graphs in this gure suggest that both the standard deviation and correlation coecient can change dramatically over the course of a few

Electronic copy available at: http://ssrn.com/abstract=2411493

years. For example, there was a ten-fold increase in the standard deviation of the stock market returns over the period 2006-08. During the same period the standard deviation of the bond market returns increased by a factor of four, whereas the correlation coecient experienced a change from virtually zero to a signicantly negative value. Therefore it is only logical to assume that in forecasting the covariance matrix one has to take into account the time-varying nature of variances and covariances. The objective of this paper is to compare the performances of dierent covariance matrix forecasting methods. The dierent forecasting methods are evaluated on a statistical basis and on a practical basis. For this purpose we perform three studies. In the rst study we directly evaluate the covariance matrix forecasting methods by comparing their out-of-sample forecast accuracy. In the second study we evaluate dierent methods by looking at their ability, in outof-sample tests, to optimize risk and improve risk-adjusted performance of minimum-variance portfolios. In the nal study we evaluate dierent methods by comparing out-of-sample performance of minimum-variance portfolios with a volatility target mechanism. Specically, we evaluate the ability of dierent covariance forecasting methods to keep the portfolio volatility at a target level and improve the portfolio risk-adjusted performance. In all our studies, employing a xed size data window of 10 years that is rolled over time, we generate out-of-sample forecasts of covariance matrix using 5 distinct methods. The rst method, which is most common in practice, uses the (rolling) sample covariance matrix as a predictor for the future covariance matrix. In other words, in this approach the equally weighted covariance matrix is estimated using a 10-year lookback period. In order to decrease the estimation error of the sample covariance matrix, in the second method we use the shrinkage estimation of the covariance matrix proposed by Ledoit and Wolf (2004). This is the application of the idea of shrinking estimation pioneered by Stein (1956). The third method of estimation of covariance matrix is popularized by RiskMetrics
TM

group and uses the expo-

nentially weighted covariance matrix. This method of forecasting is usually denoted as the Exponentially Weighted Moving Average (EWMA) model. The other two methods of covariance matrix forecasting employ multivariate Generalized AutoRegressive Conditional Heteroskedasticity (GARCH) models. The univariate GARCH modeling started with the seminal paper by Bollerslev (1986) and proved to be indispensable in modeling and forecasting time-varying nancial asset volatility. Unfortunately, a direct 3

Electronic copy available at: http://ssrn.com/abstract=2411493

0.15

0.10

0.05

0.00 1995 2000 2005 2010

Panel A: Standard deviation of USA Govt. Bond Index

0.25

0.20

0.15

0.10

0.05

0.00 1995 2000 2005 2010

Panel B: Standard deviation of MSCI USA Stock Index

0.5

0.0

0.5

1995

2000

2005

2010

Panel C: Correlation between MSCI USA Stock Index and USA Govt. Bond Index Figure 1: Panels A and B plot the monthly realized standard deviations of returns on the US stock market and bond market indices. Panel C plots the monthly realized correlation coecient between the returns on these indices. The standard deviations and correlation coecient are computed using daily returns.

extension of univariate GARCH models to multivariate GARCH models (examples are the VEC-GARCH model of Bollerslev, Engle, and Wooldridge (1988) and the BEKK-GARCH model dened in Engle and Kroner (1995)) suer from the curse of dimensionality and cannot be used to estimate covariance matrices of many nancial assets. With a few nancial assets, however, Pojarliev and Polasek (2001) and Pojarliev and Polasek (2003) demonstrate that portfolios based on BEKK-GARCH covariance matrix forecasts outperform portfolios based on sample covariance matrix forecasts. In our study we focus on relatively recent multivariate GARCH models that can be applied to estimate large covariance matrices. These models are the DCC-GARCH model of Engle (2002) and the GO-GARCH model of van der Weide (2002). To ensure that our ndings are not dataset-specic, in our studies we use 9 empirical datasets. To guarantee that our results are not conned to some particular historical episode, we estimate out-of-sample performances of dierent covariance forecasting models over a rather long period, from January 1995 to December 2013, that covers a series of alternating turbulent and calm stock market times. Our main ndings can be summarized as follows. We nd that shrinkage of sample covariance matrix improves neither the forecast accuracy nor the performance of minimum-variance portfolios. In contrast, multivariate GARCH models provide superior covariance matrix forecasts compared to the forecast based on the rolling sample covariance matrix. Specically, switching from the sample covariance matrix forecast to a multivariate GARCH forecast allows one to reduce the forecasting error and portfolio tracking error by more than 50%. At the same time, using a multivariate GARCH model for portfolio construction leads to a better risk-adjusted performance of optimized portfolios. Specically, in case the minimum-variance portfolio is implemented with a volatility target mechanism, the portfolio Sharpe ratio can be increased up to 50%. We also nd that the simple EWMA covariance matrix forecast performs only slightly worse than the multivariate GARCH forecast. Surprisingly, the minimum-variance portfolios constructed on the basis of the EWMA covariance matrix forecast generally show the best risk-adjusted performance. The rest of the paper is organized as follows. Section 2 describe our data, the construction of minimum-variance portfolios, and dierent covariance matrix forecasting methods. Section 3 presents the results of our empirical studies, whereas Section 4 draws the conclusions.

2
2.1

Data and Methodology

Data

In our study we use nine empirical datasets that are listed in Table 1. These datasets are similar to the datasets used in the previous studies by DeMiguel et al. (2009) and Kritzman et al. (2010). The rst eight datasets come from the data library of Kenneth French.1 All of them represent value-weighted portfolios formed using dierent criteria. For example, the dataset of portfolios formed on size consists of returns on 10 stock portfolios sorted by market equity. The portfolios are constructed at the end of each June using the June market equity and NYSE breakpoints. As another example, the dataset of industry portfolios consists of returns on 10 industry portfolios in the US. The 10 industries considered are Consumer Non-Durables, Consumer Durables, Manufacturing, Energy, High-Tech, Telecommunication, Wholesale and Retail, Health, Utilities, and Others.

# 1 2 3 4 5 6 7 8 9

Dataset Portfolios formed on Portfolios formed on Portfolios formed on Portfolios formed on Industry portfolios Portfolios formed on Portfolios formed on Portfolios formed on Core asset classes

size book-to-market momentum size and book-to-market size and momentum size and short-term reversal size and long-term reversal

N 10 10 10 6 10 10 10 10 7

Table 1: List of datasets considered. N denotes the number of portfolios in each dataset. The ninth dataset covers the seven major asset classes and is obtained from the Thomson Reuters Datastream database. The asset classes in this dataset are listed in Table 2. The data for all nine datasets come at the daily frequency and cover the period from January 1st, 1986 to December 31st, 2013. In addition we use the 90-day nominal US TBill rate as a proxy for the risk-free rate of return. The daily TBill rate is also obtained from the data library of Kenneth French.
1

http://mba.tuck.dartmouth.edu/pages/faculty/ken.french/data library.html

# 1 2 3 4 5 6 7

Asset class US Benchmark 30 Year Govt. Bond Index MSCI EUROPE Stock Index MSCI Emerging Markets Stock Index MSCI PACIFIC Stock Index MSCI USA Stock Index S&P USA REIT Index S&P Goldman Sachs Commodity Index

Mnemonic BMUS30Y MSEROP MSEMKF MSPACF MSUSAML SBBRUSL GSCITOT

Table 2: List of core asset classes and their mnemonics in the Thomson Reuters Datastream database.

2.2

Minimum Variance Portfolios

We assume that the investment universe consists of n risky assets and one risk-free asset. We denote by r t the n 1 vector of asset returns at time t and by wt the n 1 vector of portfolio weights at time t such that the risky portfolio return and volatility at time t is given by
rP t = r t wt ,

P t =

, wt t wt

where t denotes the n n covariance matrix of asset returns for time t. Even though the data for our study come at the daily frequency, we re-estimate the covariance matrix and re-balance the composition of the risky portfolio at the month-end only. The t the risky portfolio in our study is the global minimum variance portfolio. We denote by covariance matrix forecast for month t made at the end of month t 1. The global minimum variance portfolio is constructed by solving numerically the following quadratic optimization problem with one equality and n inequality constraints
t wt wt ,

min
wt

subject to

wt 1 = 1

and

wt 0,

where 1 and 0 denote the n 1 vectors of ones and zeros respectively. Note that the constraint wt 0 implies that the global minimum variance portfolio is long-only portfolio, that is, short sales are prohibited. The capital allocation consists of investing proportion at in the risky portfolio and, consequently, 1 at in the risk-free asset which provides the return of rf t at time t. Thus, the

return on the complete portfolio at time t is given by rCt = at rP t + (1 at )rf t .

We consider two distinct cases. In the rst case at = 1, that is, all money is allocated to the risky portfolio only. In the second case the value of at is chosen to keep the volatility of the complete portfolio at a target level denoted by target ( at = min ) target max ,a , P t

where amax is the maximum allowable exposure to the risky portfolio.

2.3

Covariance Matrix Forecasting Methods

We assume that the vector of daily asset returns is given by

r t = + t , where is the n 1 vector of asset mean returns, and t is the n 1 vector of random disturbances at time t such that E [t ] = 0. The covariance matrix of asset returns at time t is the n n matrix t = Cov [t ]. In order to compute the composition of the global minimum variance portfolio for time t, the covariance matrix of asset returns for time t needs to be forecasted using the information available at time t 1. Below we review the covariance matrix forecasting methods employed in this study.

2.3.1

Rolling Historical Covariance

In this method, the covariance matrix of asset returns for time t is forecasted using an equallyweighted moving average calculated on a xed size data window that is rolled over time. Denoting the length of this window by L (also called the length of the lookback period), the covariance matrix for time t is forecasted using
t1 1 Hist = i t i. L i=tL

Since our data come at the daily frequency, but the composition of the global minimum variance portfolio is revised at the month-end, we actually need to forecast the covariance matrix for the subsequent month. This is done by multiplying each element of the daily covariance matrix by the number of days in the subsequent month.

2.3.2

Rolling Historical Covariance with Shrinkage

The shrinkage estimation of the covariance matrix is designed to decrease the error in estimation by using the estimator of the form Target Shrink Hist , = (1 t ) + t t t t Hist Target where 0 < t < 1, is the historical covariance over rolling window of length L, and t t is the shrinkage target which is also computed over rolling window of length L. In essence, this estimator shrinks the covariance matrix toward the shrinkage target. In our study, we use the shrinkage estimator proposed by Ledoit and Wolf (2004), who take the shrinkage target to be the covariance matrix in the constant correlation model.

2.3.3

Exponentially Weighted Moving Average

TM

This approach to forecast the covariance matrix is popularized by RiskMetrics

group. In

this approach the exponentially weighted covariance matrix is estimated using the following recursive form EWMA EWMA , = (1 )t1 t t1 + t1 where 0 < < 1 is the decay constant. We follow the recommendations of RiskMetrics
TM

group and estimate the daily EWMA covariance matrix with = 0.97. The monthly EWMA covariance matrix is obtained by multiplying each element of the daily EWMA covariance matrix by the number of days in the subsequent month.

2.3.4

Dynamic Conditional Correlation GARCH

The Dynamic Conditional Correlation (DCC) GARCH model belongs to the family of multivariate GARCH models. In the DCC-GARCH model the covariance matrix is decomposed

into the matrix of standard deviations and the correlation matrix as

DCC-GARCH = Dt Rt Dt , t

where Dt = diag(1t , . . . , nt ) is the diagonal matrix containing the standard deviations of asset returns, and Rt = [ijt ] is the correlation matrix. The elements of Dt are modelled as univariate GARCH processes. In our study we use the DCC-GARCH(1,1) where the variances of each asset return are modelled as GARCH(1,1) which is given by
2 2 2 it = 0i + 1i it 1 + 1i it1 .

(1)

The standardized residuals from the estimation of the GARCH(1,1) models are used to estimate the DCC-GARCH(1,1) correlation specication
1 1 Rt = Q Qt Q , t t

where Qt = (1 a b)Q + aet1 e t1 + bQt1 ,

1 et = D t t are standardized errors, Q = E [et et ] is the unconditional covariance matrix of

standardized errors, and Q is a diagonal matrix with the square root of the diagonal elements of Qt at the diagonal, Q = (diag(Qt )) 2 . We estimate the DCC-GARCH(1,1) model using the daily data and rolling window of L observations. Then the estimated parameters are used to perform the N -day ahead forecast of the (daily) covariance matrix where N is the number of days in the subsequent month. The monthly covariance matrix is obtained by summing N daily covariance matrixes.
1

2.3.5

Generalized Orthogonal GARCH

The Generalized Orthogonal (GO) GARCH model also represents a multivariate generalization of a univariate GARCH model. In the GO-GARCH model the vector of random disturbances t is modelled as a linear combination of n unobserved independent factors ft

t = Zft , 10

where Z is the n n invertible matrix that links the unobserved components with the observed variables. The unobservable components are normalized to have unit variance such that the unconditional covariance matrix of t
= E [t t ] = ZZ .

The independent factors ft are modelled as univariate GARCH. In our study we use the GOGARCH(1,1) where the factors are modelled by a GARCH(1,1) process Ht = diag(h1t , . . . , hnt )
2 2 h2 it = 0i + 1i hit1 + 1i fit1 .

(2)

The conditional covariance matrix of t is given by GO-GARCH = ZHt Z . t Denote by P the n n matrix that contains the orthogonal elements of and by the n n diagonal matrix that contains the corresponding eigenvalues, = diag(1 , . . . , n ). These 1 t1 matrixes can be estimated since the matrix can be estimated consistently by L i=tL i i . The singular value decomposition of Z yields Z = P 2 U ,
1

where U is the n n orthogonal matrix that can be identied from the structure of the conditional covariance matrix t . Similarly to the DCC-GARCH model, we estimate the GO-GARCH(1,1) model using the daily data and rolling window of L observations. Then the estimated parameters are used to perform the N -day ahead forecast of the (daily) covariance matrix where N is the number of days in the subsequent month. The monthly covariance matrix is obtained by summing N daily covariance matrixes.

11

Empirical Results

In this section we evaluate dierent covariance matrix forecasting methods by performing three studies. In the rst study we directly evaluate the covariance matrix forecasting methods by comparing their forecast accuracy. In the second and the third study we evaluate the covariance matrix forecasting methods in practical settings. In particular, we compare the performances of dierent methods by looking at their ability to optimize risk and improve risk-adjusted performance of minimum variance portfolios and the minimum variance portfolios with a volatility target. In all our studies we perform out-of-sample forecast of the monthly covariance matrix. That is, the covariance matrix for month t is forecasted on the basis of information available at the end of month t 1. More specically, our forecasts are based on the rolling-window estimation scheme by using the lookback period of L = 120 months. With this choice, our initial in-sample period is from January 1st, 1986 to December, 31st 1994 and, consequently, the performances of dierent forecasting methods are evaluated over the period from January 1st, 1995 to December 31st, 2013.

3.1

Covariance Matrix Forecast Accuracy

In this study we evaluate the covariance matrix forecast accuracy provided by each distinct method of forecasting. Let t = [ijt ] denote the monthly realized covariance matrix for month t = [ t and ijt ] denote a forecast of t made at the end of month t 1. Both the matrixes are obtained using the daily returns. The month t squared forecast error for single covariance is given by (ijt ijt )2 . To avoid double accounting of covariance forecast errors, we compute the month t total Squared Forecast Error (SFE) for covariance matrix t by summing up the squared forecast errors for single covariances in the upper-triangular part of matrix t including the main diagonal
n i i=1 j =1

SF Et =

(ijt ijt )2 .

The Mean Squared Forecast Error (MSFE) is computed as

M 1 SF Et , M t=1

M SF E =

12

0.0100

Mean Squared Forecasting Error

0.0075

0.0050

0.0025

0.0000 HIST SHRINK EWMA DCCGARCH GOGARCH

Figure 2: Average forecast accuracy over the period 1995-2013 versus the method of covariance matrix forecasting. For each distinct method of forecasting, this gure plots the average of Mean Squared Forecast Errors over all datasets. HIST denotes the rolling historical covariance; SHRINK denotes the rolling historical covariance with shrinkage; EWMA denotes the exponentially weighted moving average covariance; DCC-GARCH denotes the Dynamic Conditional Correlation GARCH; GO-GARCH denotes the Generalized Orthogonal GARCH. In all methods the length of the rolling estimation window amounts to 120 months. where M is the number of months in the out-of-sample period. Table 3 reports the MSFE produced by dierent covariance matrix forecasting methods for each dataset separately and the average MSFEs over all datasets. Figure 2 plots the average of MSFEs over all datasets for dierent forecasting methods. The results reported in Table 3 advocate that both the rolling historical covariance method and the rolling historical covariance with shrinkage method provide identical forecast accuracy. That is, shrinkage does not produce better forecast of the covariance matrix for the subsequent month than the rolling historical covariance. In contrast, as compared with the two rst methods of forecasting, the remaining three methods allow to reduce the MSFE by approximately 50%. The dierences in forecast accuracy between these three methods are marginal. In particular, the methods based on the multivariate GARCH models produce slightly better forecasts than the method that uses EWMA, and the forecast produced by the DCC-GARCH model is marginally better than that of the GO-GARCH model.

3.2

Performance of Minimum Variance Portfolios

In this study we evaluate the out-of-sample performance of minimum variance portfolios where the composition of the (global minimum variance) portfolio for month t is determined on 13

Dataset

14

Portfolios formed on size Portfolios formed on book-to-market Portfolios formed on momentum Portfolios formed on size and book-to-market Industry portfolios Portfolios formed on size and momentum Portfolios formed on size and short-term reversal Portfolios formed on size and long-term reversal Core assets Averages over all datasets

HIST 0.0069 0.0080 0.0149 0.0081 0.0081 0.0149 0.0079 0.0153 0.0065 0.0101

Covariance matrix forecasting method SHRINK EWMA DCC-GARCH GO-GARCH 0.0069 0.0040 0.0033 0.0028 0.0080 0.0042 0.0035 0.0043 0.0149 0.0069 0.0061 0.0070 0.0081 0.0042 0.0036 0.0040 0.0081 0.0052 0.0045 0.0043 0.0149 0.0069 0.0061 0.0070 0.0079 0.0048 0.0040 0.0039 0.0153 0.0093 0.0081 0.0074 0.0065 0.0032 0.0023 0.0032 0.0101 0.0054 0.0046 0.0049

Table 3: Forecast accuracy over the period 1995-2013 versus the method of covariance matrix forecasting. This table reports the Mean Squared Forecast Error (MSFE) produced by dierent covariance matrix forecasting methods. HIST denotes the rolling historical covariance; SHRINK denotes the rolling historical covariance with shrinkage; EWMA denotes the exponentially weighted moving average covariance; DCC-GARCH denotes the Dynamic Conditional Correlation GARCH; GO-GARCH denotes the Generalized Orthogonal GARCH. In all methods the length of the rolling estimation window amounts to 120 months.

the basis of the covariance matrix forecast provided at the end of month t 1. Given the daily returns generated by the minimum variance portfolio for each dataset and each distinct covariance forecasting method, we compute two measures of performance. The rst measure is the Mean Tracking Error (MTE) which is computed as follows. Denote by w t the ex-ante vector of weights of the global minimum variance portfolio for month t and by wt the ex-post vector of weights of the global minimum variance portfolio for month t. The former one is determined t provided by each distinct method of forecasting at the end of month t 1, on the basis of and the latter one is determined on the basis of t which is the realized covariance matrix for
oos the out-of-sample (i.e., realized) monthly standard deviation month t. Further denote by P t is the in-sample monthly standard deviation of of the minimum variance portfolio and by P t

the minimum variance portfolio. These standard deviations are given by

, w t t w t

oos P t =

is P t =

. wt t wt

The month t Tracking Error (TE) is dened as the squared dierence between the in-sample and out-of-sample standard deviations of the minimum variance portfolio ) ( is oos 2 . T Et = P t P t

The MTE is computed as MTE =

M 1 T Et , M t=1

where M is the number of months in the out-of-sample period. The second measure of performance is the (out-of-sample) Sharpe ratio of the minimum variance portfolio. Table 4 reports the out-of-sample performance of minimum variance portfolios over the period 1995-2013 versus the method of covariance matrix forecasting. Specically, this table reports the portfolio MTE and the Sharpe ratio produced by dierent covariance matrix forecasting methods. Panel A in Figure 3 plots the average of MTEs over all datasets whereas Panel B in the same gure plots the average of Sharpe ratios over all datasets. The results reported in Table 4 suggest that both the rolling historical covariance method and the rolling historical covariance with shrinkage method provide virtually identical MTEs and Sharpe ratios. That is, shrinkage does not produce better performance of minimum variance portfolios 15

Statistics

HIST

Covariance matrix forecasting method SHRINK EWMA DCC-GARCH GO-GARCH 1.22e-05 0.61 2.35e-05 0.77 4.98e-05 0.58 2.46e-05 0.78 2.68e-05 0.64 4.98e-05 0.58 3.08e-05 0.63 1.28e-04 0.49 2.14e-05 1.19 4.08e-05 0.70 2.16e-05 0.61 2.43e-05 0.72 4.23e-05 0.53 1.86e-05 0.71 3.24e-05 0.69 4.23e-05 0.53 3.66e-05 0.66 7.27e-05 0.56 2.01e-05 1.15 3.45e-05 0.68 1.90e-05 0.61 2.80e-05 0.71 8.50e-05 0.52 1.69e-05 0.76 3.25e-05 0.73 8.50e-05 0.52 5.00e-05 0.69 8.13e-05 0.51 2.87e-05 0.99 4.74e-05 0.67

Portfolios formed on size Mean tracking error 5.37e-05 7.23e-05 Sharpe ratio 0.52 0.51 Portfolios formed on book-to-market Mean tracking error 1.18e-04 1.15e-04 Sharpe ratio 0.66 0.66 Portfolios formed on momentum Mean tracking error 1.24e-04 1.23e-04 Sharpe ratio 0.51 0.54 Portfolios formed on size and book-to-market Mean tracking error 1.16e-04 1.16e-04 Sharpe ratio 0.71 0.70 Industry portfolios Mean tracking error 6.27e-05 6.21e-05 Sharpe ratio 0.71 0.72 Portfolios formed on size and momentum Mean tracking error 1.24e-04 1.23e-04 Sharpe ratio 0.51 0.54 Portfolios formed on size and short-term reversal Mean tracking error 5.34e-05 5.45e-05 Sharpe ratio 0.76 0.75 Portfolios formed on size and long-term reversal Mean tracking error 1.21e-04 1.18e-04 Sharpe ratio 0.51 0.53 Core assets Mean tracking error 4.07e-05 4.09e-05 Sharpe ratio 0.74 0.74 Averages over all datasets Mean tracking error 9.03e-05 9.17e-05 Sharpe ratio 0.63 0.63

Table 4: Out-of-sample performance of minimum variance portfolios over the period 19952013 versus the method of covariance matrix forecasting. This table reports the portfolio Mean Tracking Error and the Sharpe ratio produced by dierent covariance matrix forecasting methods. HIST denotes the rolling historical covariance; SHRINK denotes the rolling historical covariance with shrinkage; EWMA denotes the exponentially weighted moving average covariance; DCC-GARCH denotes the Dynamic Conditional Correlation GARCH; GO-GARCH denotes the Generalized Orthogonal GARCH. In all methods the length of the rolling estimation window amounts to 120 months.

16

0.6 7.5e05

Mean tracking error

Sharpe ratio
HIST SHRINK EWMA DCCGARCH GOGARCH

0.4

5.0e05

2.5e05

0.2

0.0e+00

0.0 HIST SHRINK EWMA DCCGARCH GOGARCH

Panel A

Panel B

Figure 3: Average out-of-sample performance of minimum variance portfolios over the period 1995-2013 versus the method of covariance matrix forecasting. For each distinct method of forecasting, Panel A plots the average of the Mean Tracking Errors over all datasets whereas Panel B plots the average of the Sharpe ratios over all datasets. HIST denotes the rolling historical covariance; SHRINK denotes the rolling historical covariance with shrinkage; EWMA denotes the exponentially weighted moving average covariance; DCC-GARCH denotes the Dynamic Conditional Correlation GARCH; GO-GARCH denotes the Generalized Orthogonal GARCH. In all methods the length of the rolling estimation window amounts to 120 months. than the rolling historical covariance. As opposed to the two rst methods of forecasting, the remaining three methods allow to reduce the average Mean Tracking Error by 50-65%. The forecast obtained using the DCC-GARCH model produces the least MTE. Somewhat surprisingly, the use of the EWMA model produces a smaller MTE than the use of the much more advanced GO-GARCH model; yet the dierences between their MTEs are not substantial. When it comes to the Sharpe ratios of the minimum variance portfolios, even though the performance of the portfolios that use either EWMA, DCC-GARCH, or GO-GARCH method of forecasting is higher than the performance of the portfolios that use the other two methods of forecasting, the dierences are marginal. In particular, the maximum average improvement in the Sharpe ratio of the minimum variance portfolio amounts to 11% increase in the value of this performance measure.

3.3

Performance of Minimum Variance Portfolios with a Volatility Target

In the last study we evaluate the out-of-sample performance of minimum variance portfolios with a volatility target mechanism. The performance of these portfolios is also measured by the

17

Mean Tracking Error and the Sharpe ratio. To implement this portfolio construction method, we rst dene the composition of the global minimum variance portfolio for month t on the basis of the covariance matrix forecast provided at the end of month t 1. The forecasted standard deviation of the minimum variance portfolio for month t is given by P t =
, tw w t t

where w t the ex-ante vector of weights of the global minimum variance portfolio for month t t is the covariance matrix forecast for month t. Then, given the volatility target target , and the global minimum variance portfolio is mixed with the risk-free asset in order to keep the volatility of the complete portfolio at the target level. Specically, the weight of the minimum variance portfolio in the complete portfolio is determined as ( at = min ) target max ,a . P t

Given the daily returns generated by the minimum variance portfolio with a volatility target, we compute the portfolio out-of-sample monthly standard deviation which is denoted by Ct . The month t Tracking Error and the Mean Tracking Error are computed as ( ) target 2
M 1 MTE = T Et , M t=1

T Et = Ct

where M is the number of months in the out-of-sample period. In our study the annual volatility target, target , is set to be 10% and the maximum risk exposure, amax , is set to be 150%. Table 5 reports the out-of-sample performance of minimum variance portfolios with a volatility target over the period 1995-2013 versus the method of covariance matrix forecasting. Specically, this table reports the portfolio MTE and the Sharpe ratio produced by dierent covariance matrix forecasting methods. Panel A in Figure 4 plots the average of MTEs over all datasets, whereas Panel B in the same gure plots the average of Sharpe ratios over all datasets. The results on the portfolio MTE reported Table 5 agree with those presented in the previous subsection. In brief, there is no dierences in the MTEs produced by the rolling historical covariance method and the rolling historical covariance with shrinkage method. The 18

Statistics

HIST

Covariance matrix forecasting method SHRINK EWMA DCC-GARCH GO-GARCH 1.83e-04 0.72 1.50e-04 0.89 1.44e-04 0.71 1.59e-04 0.93 1.39e-04 0.80 1.44e-04 0.71 1.48e-04 0.82 1.43e-04 0.66 8.15e-05 1.45 1.43e-04 0.85 1.59e-04 0.75 1.28e-04 0.84 1.26e-04 0.66 1.48e-04 0.88 1.19e-04 0.81 1.26e-04 0.66 1.28e-04 0.85 1.34e-04 0.70 6.13e-05 1.37 1.25e-04 0.84 1.62e-04 0.73 1.22e-04 0.82 1.20e-04 0.65 1.50e-04 0.89 1.18e-04 0.90 1.20e-04 0.65 1.20e-04 0.84 1.25e-04 0.65 7.20e-05 1.17 1.23e-04 0.81

Portfolios formed on size Mean tracking error 3.61e-04 3.66e-04 0.46 0.45 Sharpe ratio Portfolios formed on book-to-market Mean tracking error 3.38e-04 3.41e-04 Sharpe ratio 0.59 0.60 Portfolios formed on momentum Mean tracking error 2.99e-04 3.01e-04 Sharpe ratio 0.44 0.47 Portfolios formed on size and book-to-market Mean tracking error 3.70e-04 3.75e-04 Sharpe ratio 0.65 0.65 Industry portfolios Mean tracking error 2.85e-04 2.83e-04 Sharpe ratio 0.63 0.64 Portfolios formed on size and momentum Mean tracking error 2.99e-04 3.01e-04 Sharpe ratio 0.44 0.47 Portfolios formed on size and short-term reversal Mean tracking error 2.80e-04 2.85e-04 0.71 0.71 Sharpe ratio Portfolios formed on size and long-term reversal Mean tracking error 2.90e-04 2.95e-04 Sharpe ratio 0.45 0.48 Core assets Mean tracking error 1.44e-04 1.43e-04 Sharpe ratio 0.75 0.76 Averages over all datasets Mean tracking error 2.96e-04 2.99e-04 Sharpe ratio 0.57 0.58

Table 5: Out-of-sample performance of minimum variance portfolios with a volatility target over the period 1995-2013 versus the method of covariance matrix forecasting. This table reports the portfolio Mean Tracking Error and the Sharpe ratio produced by dierent covariance matrix forecasting methods. The annualized target volatility of minimum variance portfolios is 10%. HIST denotes the rolling historical covariance; SHRINK denotes the rolling historical covariance with shrinkage; EWMA denotes the exponentially weighted moving average covariance; DCC-GARCH denotes the Dynamic Conditional Correlation GARCH; GO-GARCH denotes the Generalized Orthogonal GARCH. In all methods the length of the rolling estimation window amounts to 120 months.

19

other three methods of forecasting allow to reduce the MTE by 50-60%. In contrast to the results presented in the previous subsection, the EWMA, DCC-GARCH, and GO-GARCH methods of covariance matrix forecasting allow not only to reduce substantially the portfolio tracking error, but at the same time these methods allow to increase considerably the portfolio risk-adjusted performance. In particular, better methods of covariance matrix forecasting allow to increase the portfolio Sharpe ratio by 40-50%. The explanation for why better forecast accuracy allows one to achieve better risk-adjusted portfolio performance lies in the fact that the contemporaneous relation between the market return and volatility is found to be negative, see French, Schwert, and Stambaugh (1987). That is, usually over the periods with above average volatility the market return is below average. Since the volatility is predictable, keeping the volatility at a target level reduces exposure to stocks when volatility is high, and increases that exposure when volatility is low. The advantage of the minimum-variance portfolio with a volatility target over the pure minimum-variance portfolio comes from two eects: a reduction in volatility and an increase in returns; both these eects contribute to a better risk-return tradeo. Yet in order to realize this advantage one needs to forecast the future volatility with high accuracy. Poor forecast accuracy results in deterioration of the performance of minimum-variance portfolios. This eect is clearly seen by comparing the Sharpe ratios of the minimum-variance portfolios and minimum-variance portfolios with a volatility target (see Tables 4 and 5) when one uses the sample covariance matrix (with and without shrinkage) as a forecast for future covariance matrix.

Conclusions

Our research showed that there are large dierences between the covariance matrix forecasting methods. For those researchers and practitioner who routinely use the sample covariance matrix as a forward-looking estimate for the future covariance matrix our empirical results imply that the out-of-sample performance of minimum-variance portfolios can be improved substantially by adopting multivariate GARCH models. We found that minimum-variance portfolios, constructed by using multivariate GARCH forecasts, deliver the best performance in terms of risk optimization and portfolio tracking error. In particular, switching from the sample covariance matrix forecast to a multivariate GARCH forecast allows one to reduce the

20

3e04 0.8

0.6

Mean tracking error

2e04

Sharpe ratio
1e04 0e+00 HIST SHRINK EWMA DCCGARCH GOGARCH

0.4

0.2

0.0 HIST SHRINK EWMA DCCGARCH GOGARCH

Panel A

Panel B

Figure 4: Average out-of-sample performances of minimum variance portfolios with a volatility target over the period 1995-2013 versus the method of covariance matrix forecasting. The annualized target volatility of minimum variance portfolios is 10%. For each distinct method of forecasting, Panel A plots the average of Mean Tracking Errors over all datasets whereas Panel B plots the average of Sharpe ratios over all datasets. HIST denotes the rolling historical covariance; SHRINK denotes the rolling historical covariance with shrinkage; EWMA denotes the exponentially weighted moving average covariance; DCC-GARCH denotes the Dynamic Conditional Correlation GARCH; GO-GARCH denotes the Generalized Orthogonal GARCH. In all methods the length of the rolling estimation window amounts to 120 months. forecasting error and portfolio tracking error by more than 50%. The risk-adjusted performance of the portfolios based on multivariate GARCH forecast is also superior compared to that of the portfolios based on sample covariance matrix forecast. The dierences in performances is especially large when minimum-variance portfolios are implemented with a volatility target mechanism; in this case the risk-adjusted performance can be improved up to 50%. We tested two multivariate GARCH models, DCC-GARCH and GO-GARCH, and in our tests the performance of the DCC-GARCH model is found to be marginally better than that of the GO-GARCH model. There is a common impression that shrinkage of sample covariance matrix improves the accuracy of covariance matrix forecast. Our tests revealed that this impression is wrong. We found that shrinkage is a computationally intensive method that improves neither the forecast accuracy nor the performance of minimum-variance portfolios. We also found that the simple EWMA covariance matrix forecast performs only slightly worse than the multivariate GARCH forecast. Therefore those researchers and practitioner who nd that multivariate GARCH

21

models are cumbersome in implementation are advised to adopt the EWMA method. This method is not only computationally simple and fast, but also generates the best risk-adjusted performance of minimum-variance portfolios in the majority of our empirical tests.

References
Bollerslev, T. (1986). Generalized Autoregressive Conditional Heteroskedasticity, Journal of Econometrics, 31 (3), 307327. Bollerslev, T., Engle, R. F., and Wooldridge, J. M. (1988). A Capital Asset Pricing Model with Time-Varying Covariances, Journal of Political Economy, 96 (1), 116131. Chan, L. K. C., Karceski, J., and Lakonishok, J. (1999). On Portfolio Optimization: Forecasting Covariances and Choosing the Risk Model, Review of Financial Studies, 12 (5), 937974. Clarke, R. G., de Silva, H., and Thorley, S. (2006). Minimum-Variance Portfolios in the U.S. Equity Market, Journal of Portfolio Management, 33 (1), 1024. DeMiguel, V., Garlappi, L., and Uppal, R. (2009). Optimal Versus Naive Diversication: How Inecient is the 1/N Portfolio Strategy?, Review of Financial Studies, 22 (5), 19151953. Duchin, R. and Levy, H. (2009). Markowitz Versus the Talmudic Portfolio Diversication Strategies, Journal of Portfolio Management, 35 (2), 7174. Engle, R. (2002). Dynamic Conditional Correlation: A Simple Class of Multivariate Generalized Autoregressive Conditional Heteroskedasticity Models, Journal of Business and Economic Statistics, 20 (3), 339350. Engle, R. F. and Kroner, K. F. (1995). Multivariate Simultaneous Generalized ARCH, Econometric Theory, 11 (1), 122150. French, K. R., Schwert, G., and Stambaugh, R. F. (1987). Expected Stock Returns and Volatility, Journal of Financial Economics, 19 (1), 329. Kritzman, M., Page, S., and Turkington, D. (2010). In Defense of Optimization: The Fallacy of 1/N, Financial Analysts Journal, 66 (2), 3139. Ledoit, O. and Wolf, M. (2004). Honey, I Shrunk the Sample Covariance Matrix, Journal of Portfolio Management, 30 (4), 110119. Pojarliev, M. and Polasek, W. (2001). Applying Multivariate Time Series Forecasts for Active Portfolio Management, Financial Markets and Portfolio Management, 15 (2), 201211.

22

Pojarliev, M. and Polasek, W. (2003). Portfolio Construction by Volatility Forecasts: Does the Covariance Structure Matter?, Financial Markets and Portfolio Management, 17 (1), 103116. Stein, C. (1956). Inadmissibility of the Usual Estimator for the Mean of a Multivariate Normal Distribution, In Proceedings of the 3rd Berkeley Symposium on Mathematical Statistics and Probability, pp. 197206. Berkeley: University of California Press. van der Weide, R. (2002). GO-GARCH: A Multivariate Generalized Orthogonal GARCH Model, Journal of Applied Econometrics, 17 (5), 549564.

23