A Test of Covariance Matrix Forecasting Methods

Valeriy Zakamulin
published in the Journal of Portfolio Management
by agreement with the publisher the main body of the paper has been removed from this
manuscript

Abstract
Providing a more accurate covariance matrix forecast can substantially improve the
performance of optimized portfolios. Using out-of-sample tests, in this paper, we evaluate
alternative covariance matrix forecasting methods by looking at (1) their forecast accuracy,
(2) their ability to track the volatility of the minimum-variance portfolio, and (3) their
ability to keep the volatility of the minimum-variance portfolio at a target level. We find
large dierences between the methods. Our results suggest that shrinkage of the sample
covariance matrix improves neither the forecast accuracy nor the performance of minimumvariance portfolios. In contrast, switching from the sample covariance matrix forecast to
a multivariate GARCH forecast reduces forecasting error and portfolio tracking error by
at least half. Our findings also reveal that the exponentially weighted covariance matrix
forecast performs only slightly worse than the multivariate GARCH forecast.
Key words: covariance matrix forecasting, minimum-variance portfolio optimization,
sample covariance, covariance shrinkage, exponentially weighted covariance, multivariate
GARCH, model comparison
JEL classification: C30, C52, G11, G17
Author bio: Professor of Finance, School of Business and Law, University of Agder,
Service Box 422, 4604 Kristiansand, Norway, Tel.: (+47) 38 14 10 39, E-mail:
[email protected]

The traditional portfolio optimization technique is based on the mean-variance portfolio

optimization framework developed by Markowitz [1952]. However, mean-variance portfolio
optimization is dicult to implement due to the challenges associated with forecasting the
mean returns. As a result, Markowitzs portfolio optimization technique has never been widely
used. Instead, most mutual funds managers focus on uncovering undervalued securities with
potentially high expected returns.
The volatile and turbulent markets of the decade of the 2000s boosted interest in portfolio
optimization techniques with a focus on portfolio risk optimization. These new portfolio op1

timization methods require only the forecast of the covariance matrix of returns, without the
need to forecast the mean returns. The most popular portfolio risk optimization techniques
are the minimum variance portfolio (Clarke, de Silva, and Thorley [2006], Clarke, de Silva,
and Thorley [2011]), the maximum diversification portfolio (Choueifaty and Coignard [2008]),
the risk parity portfolio (Maillard, Roncalli, and Teiletche [2010], Chaves, Hsu, Li, and Shakernia [2011], Asness, Frazzini, and Pedersen [2012]), and a portfolio with volatility targeting
(Busse [1999], Collie, Sylvanus, and Thomas [2011], Butler and Philbrick [2012], Albeverio,
Steblovskaya, and Wallbaum [2013]). All of these portfolio risk optimization techniques were
extensively back-tested using historical data and showed superior performance compared to
that of mean-variance portfolios and equally weighted portfolios.
Given the increasing popularity of portfolio risk optimization techniques, in this paper,
we focus on forecasting the covariance matrix of returns. There is a common consensus that
mean returns are notoriously dicult to forecast, whereas the covariance matrix can be easily
forecasted 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 the future covariance matrix (Chan, Karceski, and Lakonishok [1999], DeMiguel,
Garlappi, and Uppal [2009], Duchin and Levy [2009], Kritzman, Page, and Turkington [2010]).
Yet, this is a valid approach only if one assumes that the covariance matrix is either constant
over time or varies very slowly over time. This assumption does not appear to hold for financial
asset returns. On the contrary, there is a large strand of the literature that demonstrates that
financial asset returns exhibit heteroskedasticity with volatility clustering. The assumption
of constant correlation between financial asset returns also appears to be violated. As a
motivation, Exhibit 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 exhibit suggest that both the standard
deviation and the correlation coecient can change dramatically over the course of a few years.
For example, there was a ten-fold increase in the standard deviation of stock market returns
over the period 2006-08. During the same period, the standard deviation of bond market
returns increased by a factor of four, whereas the correlation coecient experienced a change
from virtually zero to a significantly negative value. Therefore, it is only logical to assume
that in forecasting the covariance matrix, one must take into account the time-varying nature
2

of variances and covariances.

The benefits promised by portfolio risk optimization depend crucially on the accuracy
of the covariance matrix forecast. Yet, there has been a shortage of studies evaluating the
performance of alternative forecasting methods. The aim of this paper is to fill this gap in the
literature by comparing the performances of dierent covariance matrix forecasting methods.
The alternative forecasting methods are evaluated on a statistical basis and on a practical
basis. For this purpose, we perform three studies. In the first study, we directly evaluate the
covariance matrix forecasting methods by comparing their out-of-sample forecast accuracy. In
the second study, we follow Chan et al. [1999], Ledoit and Wolf [2003], and Disatnik and
Benninga [2007] and evaluate dierent methods by examining their ability in out-of-sample
tests to minimize the variance of portfolios. In the final study, we evaluate the performance of
alternative covariance matrix forecasting methods by their ability in out-of-sample tests to keep
the variance of the minimum-variance portfolio at a pre-specified target level. The performance
criterion in the first study is the mean squared forecasting error, whereas in the second and
the third studies, we use the mean squared tracking error as the performance criterion.
In all of our studies, employing a fixed size data window of 10 years that is rolled over time,
we generate out-of-sample forecasts of the covariance matrix using 5 distinct methods. The
first 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. 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 method applies
the concept of shrinking estimation pioneered by Stein [1956]. The third method of estimating
TM

the covariance matrix is popularized by the RiskMetrics

group and uses the exponentially

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

0.15

0.10

0.05

0.00
1995

2000

2005

2010

Panel A: Standard deviation of US 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 US Stock Index

0.5

0.0

0.5

1995

2000

2005

2010

Panel C: Correlation between MSCI US Stock Index and USA Govt. Bond Index
Exhibit 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.

model of Bollerslev, Engle, and Wooldridge [1988] and the BEKK-GARCH model defined in
Engle and Kroner [1995]) suers from the curse of dimensionality and cannot be used to estimate covariance matrices of many financial assets. With a few financial 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 the 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 findings are not dataset-specific, in our studies, we use 9 empirical
datasets. To guarantee that our results are not confined 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; this time span covers a series of alternating
turbulent and calm stock market times.
Our main findings can be summarized as follows. We find that shrinkage of the sample
covariance matrix reduces neither the forecasting error nor the tracking error of minimumvariance portfolios. In contrast, multivariate GARCH models provide superior covariance
matrix forecasts compared to the forecast based on the rolling sample covariance matrix.
Specifically, 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%. We also find that the simple EWMA covariance matrix forecast performs only slightly
worse than the multivariate GARCH forecast.
The rest of the paper is organized as follows. The second section describes our data, the
construction of minimum-variance portfolios, and alternative covariance matrix forecasting
methods. The third section presents the results of our empirical studies, and the final section
draws conclusions.

Acknowledgements
The author would like to thank Steen Koekebakker, an anonymous referee, and the participants
at the 2014 Paris Financial Management Conference for their insightful comments. The author
would also like to thank Alexios Ghalanos for developing the R package rmgarch, which

was used to estimate the multivariate GARCH models and perform forecasting. The usual
disclaimer applies.

References
Albeverio, S., V. Steblovskaya, and K. Wallbaum. Investment Instruments With Volatility
Target Mechanism. Quantitative Finance, Vol. 13, No. 10 (2013), pp. 15191528.
Asness, C. S., A. Frazzini, and L. H. Pedersen. Leverage Aversion and Risk Parity. Financial
Analysts Journal, Vol. 68, No. 1 (2012), pp. 4759.
Bollerslev, T. Generalized Autoregressive Conditional Heteroskedasticity. Journal of Econometrics, Vol. 31, No. 3 (1986), pp. 307327.
Bollerslev, T., R. F. Engle, and J. M. Wooldridge. A Capital Asset Pricing Model with TimeVarying Covariances. Journal of Political Economy, Vol. 96, No. 1 (1988), pp. 116131.
Busse, J. A. Volatility Timing in Mutual Funds: Evidence from Daily Returns. Review of
Financial Studies, Vol. 12, No. 5 (1999), pp. 10091041.
Butler, A. and M. Philbrick. Volatility Management for Better Absolute and Risk-Adjusted
Performance. White paper, Macquarie Private Wealth Inc., 2012.
Chan, L. K. C., J. Karceski, and J. Lakonishok. On Portfolio Optimization: Forecasting
Covariances and Choosing the Risk Model. Review of Financial Studies, Vol. 12, No.5 (1999),
pp. 937974.
Chaves, D., J. Hsu, F. Li, and O. Shakernia. Risk Parity Portfolio vs. Other Asset Allocation
Heuristic Portfolios. Journal of Investing, Vol. 20, No. 1 (2011), pp. 108108.
Choueifaty, Y. and Y. Coignard. Toward Maximum Diversification. Journal of Portfolio
Management, Vol. 35, No. 1 (2008), pp. 4051.
Clarke, R. G., H. de Silva, and S. Thorley. Minimum-Variance Portfolios in the U.S. Equity
Market. Journal of Portfolio Management, Vol. 33, No. 1 (2006), pp. 1024.
Clarke, R. G., H. de Silva, and S. Thorley. Minimum-Variance Portfolio Composition.
Journal of Portfolio Management, Vol. 37, No. 2 (2011), pp. 3145.
Collie, R., M. Sylvanus, and M. Thomas. Volatility-Responsive Asset Allocation. White
paper, Russell Investments, 2011.
DeMiguel, V., L. Garlappi, and R. Uppal. Optimal Versus Naive Diversification: How Inefficient is the 1/N Portfolio Strategy? Review of Financial Studies, Vol. 22, No. 5 (2009), pp.
19151953.

Disatnik, D. J. and S. Benninga. Shrinking the Covariance Matrix. Journal of Portfolio

Management, Vol. 33, No. 4 (2007), pp. 5563.
Duchin, R. and H. Levy. Markowitz Versus the Talmudic Portfolio Diversification Strategies.
Journal of Portfolio Management, Vol. 35, No. 2 (2009), pp. 7174.
Engle, R. Dynamic Conditional Correlation: A Simple Class of Multivariate Generalized
Autoregressive Conditional Heteroskedasticity Models. Journal of Business and Economic
Statistics, Vol. 20, No. 3 (2002), pp. 339350.
Engle, R. F. and K. F. Kroner. Multivariate Simultaneous Generalized ARCH. Econometric
Theory, Vol. 11, No. 1 (1995), pp. 122150.
Kritzman, M., S. Page, and D. Turkington. In Defense of Optimization: The Fallacy of
1/N. Financial Analysts Journal, Vol. 66, No. 2 (2010), pp. 3139.
Ledoit, O. and M. Wolf. Honey, I Shrunk the Sample Covariance Matrix. Journal of Portfolio
Management, Vol. 30, No. 4 (2004), pp. 110119.
Ledoit, O. and M. Wolf. Improved Estimation of the Covariance Matrix of Stock Returns
with an Application to Portfolio Selection. Journal of Empirical Finance, Vol. 10, No. 5
(2003), pp. 603 621.
Maillard, S., T. Roncalli, and J. Teiletche. The Properties of Equally Weighted Risk Contribution Portfolios. Journal of Portfolio Management, Vol. 36, No, 4 (2010), pp. 6070.
Markowitz, H. Portfolio Selection. Journal of Finance, Vol. 7, No. 1 (1952), pp. 7791.
Pojarliev, M. and W. Polasek. Applying Multivariate Time Series Forecasts for Active Portfolio Management. Financial Markets and Portfolio Management, Vol. 15, No. 2 (2001), pp.
201211.
Pojarliev, M. and W. Polasek. Portfolio Construction by Volatility Forecasts: Does the
Covariance Structure Matter? Financial Markets and Portfolio Management, Vol. 17, No. 1
(2003), pp. 103116.
Stein, C. 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, 1956.
van der Weide, R. GO-GARCH: A Multivariate Generalized Orthogonal GARCH Model.
Journal of Applied Econometrics, Vol. 17, No. 5 (2002), pp. 549564.