You are on page 1of 9

Economic Modelling 52 (2016) 690–698

Contents lists available at ScienceDirect

Economic Modelling

journal homepage: www.elsevier.com/locate/ecmod

Hedging performance of REIT index futures: A comparison of alternative


hedge ratio estimation methods
Jian Zhou ⁎
Real Estate and Housing, College of Business and Economics, University of Guelph, 50 Stone Rd. E., Guelph, ON N1G 2 W1, Canada

a r t i c l e i n f o a b s t r a c t

Article history: Futures contracts based on REIT market indices remain an under-researched topic, given their short history. This
Accepted 5 October 2015 paper extends the literature by examining what hedge-ratio estimation method yields the most effective hedging
Available online 23 October 2015 performance of REIT futures. We include a wide range of commonly used methods and apply them to all four
global markets which have developed REIT index futures (i.e., Australia, Europe, Japan & the U.S.). By adopting
Keywords:
an out-of-sample analytical framework, our results show that there exist multiple methods in each market
REIT index futures
Hedge ratio
that can be considered best performers and the mix of best performers varies across markets. Furthermore,
Hedging performance our results suggest that constant hedge-ratio methods are not necessarily inferior to their time-varying counter-
Out-of-sample analysis parts, and that a more complicated GARCH model does not necessarily lead to better performance than a more
parsimonious one. Finally, only DCC and BEKK are found to rank consistently among the best performers across
all four markets when we examine collectively the results using different out-of-sample periods. However, this
does not mean that hedgers will always want to use them.
© 2015 Elsevier B.V. All rights reserved.

1. Introduction literature), REIT index futures have remained a fairly under-


researched subject due to their short history but have attracted increas-
Trading of stock-index-based futures contracts began in early 1980s. ing attention from researchers recently. For example, Lee and Lee
They were an immediate success. One reason for this success was that (2012) investigated the hedging effectiveness of REIT futures in
index futures have greatly extended the range of investment and risk Australia and Japan and found that the hedging is effective. Shi and Xu
management strategies available to investors. The success of stock (2013) explored the European market, and found that the spot market
index futures has since led to a proliferation of new futures and options leads the futures market in price discovery. Also focusing on the
markets tied to various indexes. In 2002 the first index futures contract European market, Lee et al. (2014) suggest that futures trading did not
based on real estate equity sector was launched in Australia. Following destabilize the spot market. Instead, it has improved the speed and
that, similar products were developed and traded at a few other ex- quality of information flowing to the spot market. They also confirm
changes around the world (i.e., the Chicago Board of Trade in 2007, the hedging effectiveness of REIT futures.
the NYSE LIFFE Euronext Paris in 2007, and the Tokyo Stock Exchange While index futures can be used as a class of speculative assets, their
in 2008). The introduction of these futures has arguably been propelled basic use is risk hedge (Kolb, 2003). For REIT index futures, Lee and Lee
by the rapid growth of the underlying listed real estate market. Accord- (2012) and Lee et al. (2014) have shown that they are effective hedging
ing to Standard & Poor Global REIT Index, the global market capitaliza- tools. However, the existing literature leaves one important question
tion of Real Estate Investment Trusts (REITs) – the dominant form of open: what is the best way to determine the hedge ratio for REIT
listed real estate investments – has grown at an average rate of 12.33% index futures? Hedge ratio refers to the number of futures contracts
per annum over the last 15 years, reaching around US$572 billion in held for each unit of the underlying asset. It is a control variable of cru-
2011. The launch of REIT index futures provide investors with an alter- cial importance in the development of effective hedging strategies.
native means of gaining exposure to this fast-expanding sector. More- Though Lee and Lee (2012) is the first to tackle the question, their inves-
over, it allows investors to hedge real estate risk of their portfolios, tigation only includes three methods. This paper aims to extend the lit-
and to enhance the liquidity of real estate investments (Newell and erature along several lines: first, we examine a much larger repertoire of
Tan, 2004; Ong and Ng, 2009). minimum-variance hedge ratio estimation methods by drawing upon
In contrast with their extensively examined stock counterparts (see the extensive econometric development. As can be seen later, the
Lien and Tse, 2002; Chen et al., 2003 for comprehensive reviews of the methods chosen cover a varying degree of complexity. We consider
those most commonly used (as surveyed in Lien and Tse, 2002 and
⁎ Tel.: +1 519 824 4120x56634; fax: +1 519 823 1964. Chen et al., 2003) and a relatively new one. Second, we make the first
E-mail address: jian@uoguelph.ca. attempt to include all four global REIT markets which have launched

http://dx.doi.org/10.1016/j.econmod.2015.10.009
0264-9993/© 2015 Elsevier B.V. All rights reserved.
J. Zhou / Economic Modelling 52 (2016) 690–698 691

index futures (i.e., Australia, Europe, Japan & the U.S.). Doing so can en- (the numerator) and variance (denominator) of Eq. (1) and then solve
hance the generality of our findings. Third, to ensure robustness of our for hedge ratio as the ratio of covariance over variance. In the finance lit-
findings, we use a variety of metrics to assess the hedging performance erature, GARCH-type models are widely used to estimate the second
of the competing methods. Finally, we examine the statistical signifi- moments of financial returns (i.e., covariance and variance). So, includ-
cance of hedging-performance comparisons. We want to see whether ed here are three popular bivariate GARCH models (CCC, DCC, & BEKK)
certain model(s) statistically outperforms the others. To this end, we and a relatively new one — copula-based GARCH. We will use the same
use Hansen et al.'s (2011) Model Confidence Set (MCS) procedure, univariate GARCH(1,1) to model variance for all four models. As such,
which is particularly suitable to compare a large number of models. To the three bivariate GARCH models distinguish each other in their way
our best knowledge, this is the first study to apply MCS to hedging to model the covariance (or equivalently, correlation) between spot
models. and future returns. In comparison with them, the copula-based
To carry out the study, we collect daily indices of closing spot GARCH allows for a more flexible modeling of the spot-futures depen-
prices and continuous futures prices for the four REIT markets. dence structure through the use of copula — a function that links togeth-
Given the forward-looking nature of hedging decision (i.e., making er univariate distributions to form a multivariate distribution. Unlike the
a decision for the next period given all currently available informa- regular bivariate GARCH, copula-based GARCH can accommodate both
tion), we focus on the out-of-sample performance of alternative linear and nonlinear dependence, and also any specific aspect of the de-
hedging methods. Our major findings are summarized as follows: pendence structure (e.g., tail dependence; see Patton, 2006; Hsu et al.,
First, statistically speaking, multiple models perform the best in 2008). It is worth noting that GARCH-related models implicitly estimate
each market and the mix of best performers varies across markets. time-varying hedge ratios, because they specify conditional (i.e., time-
Second, our results suggest that constant hedge-ratio methods are varying) second moments of financial returns.
not necessarily inferior. Though the time-varying methods appear As mentioned earlier, we focus on out-of-sample hedging perfor-
more theoretically appealing, we do find that constant methods mance. This implies that we split the sample into two periods: in-
often belong to the cohort of best performers. Third, a more compli- sample and out-of-sample. We use the in-sample period to estimate
cated model like C-GARCH does not necessarily lead to better perfor- the models and then evaluate the models' performances over the out-
mance than a more parsimonious one (i.e., DCC- or BEKK-GARCH). of-sample period. In the case of time-varying hedge ratios, the out-of-
Finally, only DCC and BEKK are ranked consistently among the best sample analysis entails forecasting the hedge ratio — a point which
performers across all four markets when results from different out- we want to emphasize. Forecasts can be done using either a recursive
of-sample periods are taken together. This testifies to the popularity window or rolling window specification (Rogoff and Stavrakeva,
of those models (e.g., Baillie and Myers, 1991; Park and Switzer, 2008). The recursive window scheme adds one more observation to
1995; Kavussanos and Nomikos, 2000; Floros and Vougas, 2006, the in-sample portion for each additional out-of-sample forecast. For
and others). However, this does not mean that hedgers will always example, if the first forecast is based on the first R observations, then
want to use them. the second forecast is based on the first R + 1 observations, etc. In con-
The remainder of this paper is organized as follows. Section 2 pro- trast, the rolling window scheme preserves the original in-sample size
vides a brief review of the competing hedge ratio estimation methods. throughout; hence, the first forecast is based on observations from 1
Section 3 discusses the data. Section 4 presents the empirical findings. to R, the second on observations from 2 to R + 1, and so on. We will con-
Section 5 concludes. sider both specifications.

2. Econometric methods 2.1. The OLS method

The objective of hedging is to offset the fluctuations in the spot posi- For a constant hedge ratio, we simply base on the initial in-sample
tion by using futures contracts. The hedger must determine a hedge data and use OLS to estimate β through the following linear regression
ratio, i.e., the ratio of futures contracts to buy or sell for each unit of model:
the spot asset. Based on the principles of portfolio theory, Ederington
(1979) and Figlewski (1984) formulate the hedger's problem and de- ΔSt ¼ α þ βΔF t þ εt : ð2Þ
rive hedge ratios that minimize the variance of the hedged portfolio.
Let S and F denote the natural logarithms of spot and futures prices, The β estimated using the initial in-sample data will be applied to
respectively. Then the optimal hedge ratio h, is derived as the whole out-of-sample period, thus making it a constant hedge
ratio method. To have time-varying hedge ratios, we apply a certain
covðΔS; ΔF Þ window scheme to Eq. (2). The hedge ratios obtained this way are
β¼ ð1Þ
varðΔ F Þ generally time-varying, as the estimation period is changing over
time.
where ΔS and ΔF represent respectively the changes in the log prices of
the spot and futures (i.e., log returns). There are a variety of methods to 2.2. The cointegration method
estimate β. Generally speaking, they can be divided into two categories:
direct estimation and indirect estimation. In what follows, we consider Since the arbitrage condition likely ties the spot and futures prices,
those most commonly used (Chen et al., 2003; Lien and Tse, 2002) they cannot drift far apart in the long run. This speaks to the possibility
and a relatively new one. that the two series could be cointegrated. If cointegration is confirmed,
Direct estimation is largely regression-based, because β, as defined we can estimate the hedge ratio using an error-correction model
in Eq. (1), can be estimated as the slope coefficient of an OLS regression (ECM):
of ΔS on ΔF. So the OLS method is a natural choice. The Cointegration
method, which considers the possible cointegration between spot and Xm Xn
ΔSt ¼ ρut−1 þ βΔ F t þ δ Δ F t−i
i¼1 i
þ θ ΔSt− j
j¼1 i
þ et ð3Þ
futures prices, can be regarded an ‘augmented’ version of OLS. Both
methods can be applied to estimate constant and time-varying versions
of hedge ratio. Under this category, we also consider the random coeffi- where ut is the error correction term from a cointegrating regression
cient method, which assumes the hedge ratio to be a random coefficient given by
and use the so-called Kalman filter to estimate time-varying β. On the
other hand, indirect estimation seeks to estimate separately covariance St ¼ a þ bF t þ ut : ð4Þ
692 J. Zhou / Economic Modelling 52 (2016) 690–698

Following the same procedure as for the OLS method, the where Q t = (qij,t) is a N × N symmetric positive definite matrix and
cointegration method can estimate both constant and time-varying meets
versions of hedge ratio.
Q t ¼ ð1−a−bÞQ þ azt−1 zt−1 0 þ bQ t−1 ð9Þ
2.3. The random coefficient method
where a and b are non-negative scalar parameters satisfying a + b b 1,
This model is particularly useful to account for the time- and Q is the unconditional covariance matrix of zt.
varying nature of the variable of interest (e.g., hedge ratio by ^ are estimated and by imposing Q = 0, we use Eq. (9) to
Once â and b 0
Bera et al., 1997 and systematic risk by Bos and Newbold, solve for the forecast of Q t + 1. Following Eq. (8), we then have Rt + 1 =
1984). Under this model, the hedge ratio is assumed to follow diag(q−1/2 −1/2 −1/2 −1/2
11,t + 1 … qNN,t + 1)Q t + 1diag(q11,t + 1 … qNN,t + 1). Finally, the one-
a random walk: step-ahead forecast of the conditional covariance matrix Ht + 1 can be
obtained as
ΔSt ¼ α þ βt ΔF t þ εt
ð5Þ
βt ¼ βt−1 þ et H tþ1 ¼ Dtþ1 Rtþ1 Dtþ1 ð10Þ

where {εt, et} is a bivariate white noise. Note that the time subscript where Dt + 1 = diag(h1/2 1/2
11,t + 1 … hNN,t + 1) is a diagonal matrix consisting of
of β now indicates that β is changing over time. It is easy to see that the forecasted conditional standard deviation of individual asset returns
Eq. (5) constitute a state space model. The hedge ratio βt, now a state based on Eq. (7). Once ht + 1 and Ht + 1 are obtained, we can get a fore-
variable in the system, can be estimated using the Kalman filter — a cast of hedge ratio based on Eq. (1).
recursive algorithm for calculating the minimum mean square
error estimate of the state variables given the appropriate informa- 2.4.3. BEKK-GARCH
tion set (Durbin and Koopman, 2001). This method only leads to The BEKK-GARCH model, proposed by Engle and Kroner (1995), es-
time-varying hedge ratios. Note that once βt is estimated, the one- timates Ht:
step-ahead hedge ratio forecast is simply Et(βt + 1) = βt, according
to Eq. (5). X
q X
p
Ht ¼ C 0 C þ Ai 0 εt−i εt−i 0 Ai þ B j 0 H t− j B j ð11Þ
i¼1 j¼1
2.4. GARCH-based models

where C, A, and B are all N × N matrices with C being upper triangular. To


Throughout the rest of the paper, we follow the convention to use Ht
ease estimation of Eq. (12), we impose restrictions on Ai & Bj by making
to denote conditional covariance (i.e., cov(ΔS, ΔF) — the numerator of
both a scalar times the identity matrix of order N (i.e., IN). Hence, in this
Eq. (1)) and ht to denote conditional variance (i.e., var(ΔF) — the
paper we estimate the following scalar BEKK-GARCH(1,1) model.
denominator of Eq. (1)).
0
H t ¼ C 0 C þ ðaI N Þ0 εt−1 εt−1 0 ðaIN Þ þ ðbIN Þ H t−1 ðbIN Þ: ð12Þ
2.4.1. CCC-GARCH
The constant conditional correlation (CCC) GARCH model, devel-
oped by Bollerslev (1990), makes the conditional covariance propor- Once model parameters (A, B, C, a b, & c) are estimated, impos-
tional to the product of the conditional standard deviations of ing some initial conditions (i.e., ε0 = 0 & H0 = 0) can yield a fore-
individual return series by assuming that the conditional correlations cast of Ht + 1 using Eq. (12). A forecast of ht + 1 can be obtained as
are time invariant. It estimates Ht as follows: before. So every element is in place for calculating a forecast of
the hedge ratio.
 qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
H t ¼ Dt RDt ¼ ρi j hii;t h j j;t ð6Þ
2.4.4. Copula-based GARCH
Copula is a function that links together univariate distributions to
where Dt = diag(h1/2 1/2
11,t … hNN,t) is a diagonal matrix whose elements are form a multivariate distribution. It has been increasingly applied in fi-
the conditional standard deviation of individual asset returns estimated nance and economics (see Patton, 2009). Consider a vector random var-
from a GARCH(1,1) model, and R = (ρij) is the matrix containing the iable, X ≡ [X1, X2, …, Xn]′, with joint distribution of F and marginal
constant conditional correlation correlations ρij. distributions F1, F2,…, Fn. Sklar's (1959) theorem states that there exists
To estimate Ht according to Eq. (6), it takes two steps: first estimate a a function C called copula which joins F to F1, F2,…, and Fn
separate GARCH(1,1) model for each return series to have hii,t:
F ðxÞ ¼ C ð F 1 ðx1 Þ; F 2 ðx2 Þ; …; F n ðxn ÞÞ: ð13Þ
hii;t ¼ wi þ αε2i;t−1 þ βhii;t−1 ; i ¼ 1; …; N ð7Þ
We can model the marginal distributions Fi's using a univariate
and then based on the estimated hii,t use a maximum likelihood criteri- GARCH model — same as we did for the regular bivariate GARCH.
on to estimate R. Once the estimated R^ ¼ ðρ^ i j Þ is obtained, we choose a What makes this model unique is the use of copula, which is a more
flexible function to characterize the dependence structure among ran-
window scheme to have a forecast of hii,t + 1, i = 1, …, N, and then a
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi dom variables. For a detailed review of copula theory, refer to Joe
^ i j hii;tþ1 h j j;tþ1 Þ. The forecast of hedge ratio easily
forecast of H tþ1 ¼ ðρ (1997) and Nelson (2006). Among the various copula forms, we follow
follows given Eq. (1). Hsu et al. (2008) to use Gaussian copula, as it is found to dominate other
copula forms in modeling direct hedge (i.e., the cash market instrument
2.4.2. DCC-GARCH being hedged is hedged by a futures contract on the same underlying in-
The dynamic conditional correlation model (DCC) GARCH of Engle strument — the situation we face here). Gaussian copula can be written
(2002) extends CCC-GARCH by making the conditional correlation as (Patton, 2006):
time dependent. The conditional correlation matrix (Rt) is assumed to (  )
Z Φ−1 ðuÞ Z Φ−1 ðvÞ
be determined as follows: 1 − r 2 −2ρrs þ s2
C ðu; vjρÞ ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi exp drds
    −∞ −∞ 2π ð1−ρ2 Þ 2ð1−ρ2 Þ
−1=2 −1=2 −1=2 −1=2
Rt ¼ diag q11;t …qNN;t Q t diag q11;t …qNN;t ð8Þ ð14Þ
J. Zhou / Economic Modelling 52 (2016) 690–698 693

Table 1 Table 2
Futures contracts specifications. Summary statistics.

Panel A: Australia Europe Mean Std. Dev. Skewness Kurtosis

Underlying index S&P/ASX 200 A-REIT Index FTSE EPRA/NAREIT Europe Australia
Index Spot −0.015 1.601 −0.647⁎ 6.946⁎
Exchange Australian Securities Exchange NYSE LIFFE Euronext Paris Futures −0.014 1.834 −0.906⁎ 31.015⁎
Currency AUD Euro
Contract size AUD$10 per index point €10 per index point Europe
Contract month Mar, Jun, Sep, Dec Mar, Jun, Sep, Dec Spot 0.001 1.449 −0.289⁎ 4.670⁎
Last trade day 3rd Thu of the expiry month 3rd Fri of the expiry month Futures 0.001 1.498 −0.505⁎ 4.356⁎

Japan
Panel B: Japan U.S. Spot 0.017 1.598 −0.400⁎ 11.250⁎
Underlying index TSE REIT Index Dow Jones U.S. Real Estate Futures 0.017 1.541 −0.299⁎ 11.073⁎
Index
U.S.
Exchange Tokyo Stock Exchange Chicago Mercantile Exchange
Spot 0.046 0.842 −0.088 2.985⁎
Currency JPY USD
Futures 0.053 0.902 −0.290⁎ 6.757⁎
Contract size ¥1000 per index point $100 per index point
Contract month Mar, Jun, Sep, Dec Mar, Jun, Sep, Dec Notes: This table shows the summary statistics for REIT spot and futures returns. The
Last trade day 2nd Fri of the expiry month 3rd Fri of the expiry month sample period of Australia, Europe, and Japan is from Jun. 16, 2008 to Dec. 31, 2014
while that of the U.S. is from Nov. 21, 2011 to Dec. 31, 2014.
⁎ Significant at 5% level.

where u and v are the marginal distributions of spot and futures Japan to Dec. 31, 2014 (a total of 1707 observations). This works well for
returns, Φ−1 is the inverse of Gaussian cumulative distribution function, all four markets except the U.S. The reason is that the U.S. REIT futures
and − 1 b ρ b 1 is the correlation coefficient. To allow time-varying cor- market crashed during the recent financial crisis and has an extended
relations, we follow Patton (2006) to specify: period of zero trading volume. It did not recover until late Nov. of
2011, forcing us to use a shorter sample period Nov. 21, 2011–Dec. 31,
0 1
2014 (812 observations) for the U.S. Given this, we should read the
~ @ωρ þ β ρ 1 X 10    
ρt ¼ Λ ρ t−1 þ α  Φ−1 ut− j Φ−1 vt− j A ð15Þ U.S. results with caution.
10 j¼1
We compute returns as the first difference in the natural logarithm
of prices. Table 2 presents the summary statistics. The average spot
~ and futures returns vary across the markets. So do the variance of the
where ΛðxÞ ≡ ð1−e−x Þð1 þ e−x Þ−1 is the modified logistic function, de-
returns (Std. Dev.). Nearly every return series (except the U.S. spot) ex-
signed to keep ρt in (−1 1) at all times, and ωρ, βρ, & α are parameters.
hibits skewness (i.e., asymmetry between negative and positive
Once the full model of Eqs. (13)–(15) is estimated, we obtain a fore-
returns). They all display significant kurtosis (i.e., fat-tailness). We
cast of ht + 1 as done before and use Eq. (15) to get a forecast of ρt. These
then test for unit root. Since our sample (except the U.S.) covers the re-
are the two ingredients needed to construct a forecast of Ht + 1. Then
cent global financial crisis period 2008–09, we opt to use break-point-
based off Eq. (1), we can easily make a forecast of the hedge ratio.
robust unit root tests. The results are presented in Panel A of Table 3.
Using the tests of Popp (2008) and Narayan and Popp (2010, 2013),
3. Data
we find that both spot and futures prices contain a unit root. This
holds for all markets under study. Based on this finding, we use the
We study the four REIT markets which have developed index futures
(Australia, Europe, Japan & the U.S.). Table 1 presents the main features
of futures contracts for those markets. As can be seen, the contracts Table 3
Unit root and cointegration test results.
share a similar trading cycle with contract months set in March, June,
September, and December. Furthermore, all of them are considered Panel A: unit-root Panel B: cointegration
mini-futures, as the contract size multiplier is pretty small (10 for Popp NP
both Australia and Europe, 1000 for Japan, and 100 for the U.S.).
Australia
For the four markets, we collect daily closing spot prices and settle- Spot −1.975 −1.961 −6.810⁎
ment prices of REIT index contracts from DataStream. However, a well- Futures −2.236 −2.276
known problem encountered in the analysis of futures contracts is that
Europe
individual contracts expire. To deal with this problem and construct a Spot −3.374 −3.254 −5.437⁎
continuous series of futures prices, we follow the practice of Alizadeh Futures −3.019 −3.027
et al. (2008): first, the nearby futures contract — the one with the
Japan
nearest delivery month to the day of trading, is specified. Prices of the Spot −2.754 −1.679 −27.141⁎
nearby futures contract are used until both of its trading volume and Futures −2.678 −2.337
open interest are exceeded by the next nearest contract.1 After that,
U.S.
prices of the next nearest contract are used.2 Among the four markets, Spot −3.439 −3.371 −9.708⁎
Japan is the latest to launch futures. So we try to create a continuous fu- Futures −3.081 −2.985
tures price series from Jun. 16, 2008 — the first day of futures trading in
Notes: Popp is the structural break unit-root test of Popp (2008), which assumes one break
while NP is the structural break unit-root test of Narayan and Popp (2010, 2013), which
1
Open interest refers to the total number of outstanding futures contracts that have not assumes two breaks. The null hypothesis of both tests is a unit root. We generate the
been settled. test statistics by assuming endogenous (unknown) break(s) and allowing for a change
2
For example, in Australia from Jun. 16–Sep. 17, 2008, the nearby contract is the one of in both the level and slope of log prices. Both tests indicate unit root as the test statistics
Sep. 2008. So the settlement prices of the Sep. 2008 contract are being used. However, on presented are well below (in terms of absolute value) the 5% critical values as shown in
Sep. 18, 2008, both the trading volume and open interest of the Sep. 2008 contract are the above two papers. Panel B shows the results of cointegration test with a break of
exceeded by the corresponding numbers of the Dec. 2008 contract. So since Sep. 18, Gregory and Hansen (1996). We use the Z⁎t test statistics and specify a break in both the
2008 we have used the settlement prices of the Dec. 2008 contract but only until Dec. level and slope of log futures and spot prices. All test statistics reject the null hypothesis
18, 2008 after which the trading volume and open interest of the Dec. 2008 are exceeded of no cointegration.
by those of the next contract — Mar. 2009. The above process then continues. ⁎ Significant at 5% level.
694 J. Zhou / Economic Modelling 52 (2016) 690–698

spot futures
Australia Europe

5 5

0 0

-5
-5

-10
2009 2010 2011 2012 2013 2014 2015 2009 2010 2011 2012 2013 2014 2015

Japan
10 5.0 U.S.

5 2.5

0
0.0

-5
-2.5

-10
-5.0

2009 2010 2011 2012 2013 2014 2015 2012 2013 2014 2015

Fig. 1. Time series plots of spot and futures returns for REITs. The sample period is Jun. 16, 2008–Dec. 31, 2014 for Australia, Europe, and Japan whereas it is Nov. 22, 2011–Dec. 31, 2014 for
the U.S.

break-robust test of Gregory and Hansen (1996) to check for variance of returns of these portfolios over the out-of-sample period is
cointegration. The results in Panel B suggest that spot and futures prices computed as
are cointegrated. This lends support to the use of the cointegration  
method. Fig. 1 presents the time-series plots of returns. Note that the se- ^ ΔF t
Var ΔSt −β ð16Þ
t
ries clearly exhibit a property of volatility clustering, where periods of
high volatility or low volatility can remain persistent for some time be- ^ are the computed hedge ratios (the time subscript would not
where β t
fore switching. Modeling such a property is the essential role of GARCH
apply for constant hedge ratios). Given Eq. (16), hedging performance
models.
can be evaluated by calculating:

Var u −Varh
4. Empirical results Variance reductionð%Þ ¼  100 ð17Þ
Varu
4.1. Hedge ratios and hedging performance
where Varu (Varh) is the variance from the unhedged (hedged) position.
Panel A of Table 4 presents the results using a recursive window for 75%
In addition to the methods discussed in Section 2, we also include a
out-of-sample. It is easy to see that hedging leads to considerable vari-
naïve strategy which is widely considered in empirical studies. This
ance reduction, which highlights the effectiveness of REIT index futures
^ ¼ 1, and is therefore a constant hedge-ratio meth-
strategy simply sets β as a risk management tool — a finding consistent with the existing liter-
od. So altogether our study considers ten methods: three constant ature. Most importantly, we find that BEKK performs the best across all
hedge ratio methods (Naïve, OLS, & Cointegration method), and seven markets, as it yields the most variance reduction. This is in contrast with
time-varying hedge ratio methods (OLS, Cointegration, Random Coeffi- the finding of Lee and Lee (2012), which reports TV-OLS as the best per-
cient, CCC-, DCC-, BEKK-, & Copula-based GARCH). Throughout the former in Australia and Japan. Clearly, their study suffers from the inclu-
paper, they are labeled respectively as Naïve, C-OLS, & C-Cointg, and sion of merely three methods and the use of a technically deficient
TV-OLS, TV-Cointg, RC, CCC, DCC, BEKK, & C-GARCH. GARCH model.3 In addition to the above result, we find that the relative
To ensure the robustness of results, we follow Phan et al. (2015) to performance of other models varies across markets. For instance, RC is
use several different choices of out-of-sample periods, namely, 75%, the 2nd best performer in Australia, Europe and the U.S. but becomes
50%, and 25% of the total sample. The first two out-of-sample evalua- one of the worst in Japan.
tions (75% & 50%) are conducted within a recursive window scheme The second metric we use is utility gain (or economic benefit). In de-
while the last one (25%) is conducted within a rolling window scheme. riving the hedge ratio of Eq. (1), we focus solely on minimizing the
We use three metrics to assess the hedging performance. The first one is
the relative variance reduction of the hedged portfolio over the un- 3
The VECH-GARCH model used by Lee and Lee (2012) does not guarantee a positive-
hedged position (e.g., Ederington, 1979). To come up with the measure, definite covariance matrix. Note that positive definiteness is a characteristic for a true co-
portfolios implied by the estimated hedge ratios are constructed and the variance matrix. See Tsay (2005) p. 448.
J. Zhou / Economic Modelling 52 (2016) 690–698 695

Table 4 A typical mean-variance utility function can be written as:


Hedging effectiveness results based on a recursive window scheme for a 75% out-of-
sample.
Et U ðxtþ1 Þ ¼ Et ðxtþ1 Þ−kVart ðxtþ1 Þ ð18Þ
Panel A Panel B Panel C

Variance Variance Utility Utility VaR VaR where k is the degree of risk aversion and xt + 1 is the return of the
reduction improvement reduction
hedged portfolio. To compute the utility, we consider an investor with
(%) (%)
medium levels of risk aversion. We assume k = 4. The expected return
Australia
and variance equal respectively the average return and variance over
Unhedged 0.878 – −3.491 – 1.563 –
Naïve 0.356 59.381 −1.432 2.059 0.977 37.487 the out-of-sample period. Panel B of Table 4 shows the results. Our
C-OLS 0.347 60.488 −1.391 2.099 0.965 38.265 focus is the utility improvement, which is computed as the difference
C-Cointg 0.346 60.561 −1.389 2.102 0.964 38.314 between the utility of a certain hedging strategy and the utility of the
TV-OLS 0.345 60.690 −1.384 2.106 0.963 38.393 unhedged one. The most salient point is that the relative ranking of
TV-Cointg 0.344 60.743 −1.382 2.108 0.962 38.431
the models under this metric is substantially consistent with that
RC 0.343 60.926 −1.372 2.118 0.963 38.360
CCC 0.347 60.486 −1.393 2.097 0.965 38.263 based on variance reduction. Also using this metric can help assess the
DCC 0.346 60.556 −1.386 2.104 0.967 38.139 impact of transaction costs. Remember that dynamic (time-varying)
BEKK 0.338 61.460 −1.352 2.138 0.958 38.709 hedging strategies require frequent rebalancing of the hedged portfolio
C-GARCH 0.348 60.398 −1.397 2.094 0.967 38.263
and are thus more costly to implement than static (constant) ones.
Europe Suppose in a certain market a dynamic strategy outperforms those static
Unhedged 1.272 – −5.069 – 1.875 – ones. An important question arises: is the utility gain high enough to off-
Naïve 0.646 49.229 −2.598 2.471 1.307 30.264
set transaction costs? Take Australia for example. BEKK — the best
C-OLS 0.569 55.220 −2.288 2.781 1.232 34.302
C-Cointg 0.575 54.795 −2.310 2.759 1.237 34.014
performing dynamic hedging method enjoys a utility gain of 0.036
TV-OLS 0.559 56.037 −2.245 2.824 1.222 34.830 (=2.138–2.102) over the C-Cointg — the best-performing static strate-
TV-Cointg 0.563 55.693 −2.263 2.805 1.226 34.613 gy. As such, hedgers would prefer BEKK over C-Cointg if 3.6% N y, where
RC 0.508 60.027 −2.047 3.022 1.160 38.125 y represents the transaction costs incurred due to the rebalancing of
CCC 0.561 55.889 −2.251 2.818 1.226 34.618
BEKK-hedged portfolio. It is well known that the transaction costs of
DCC 0.540 57.528 −2.171 2.897 1.199 36.063
BEKK 0.503 60.434 −2.025 3.044 1.155 38.398 index futures are low, typically in the range of 0.01–0.015% (Park and
C-GARCH 0.582 54.189 −2.339 2.730 1.248 33.441 Switzer, 1995). Given this, hedgers would still pick BEKK even after
transaction costs are taken into account. Following this logic, we find
Japan
Unhedged 1.242 – −4.913 – 1.890 – that BEKK also generate utility gain sufficient to compensate transaction
Naïve 0.390 68.602 −1.560 3.352 1.027 45.643 costs, compared with the respective best static strategy in all other three
C-OLS 0.370 70.173 −1.479 3.433 1.004 46.867 markets. This finding justifies the implementability of the dynamic
C-Cointg 0.391 68.496 −1.565 3.347 1.029 45.559
strategy.
TV-OLS 0.363 70.739 −1.450 3.463 0.996 47.308
TV-Cointg 0.384 69.095 −1.534 3.378 1.020 46.024
The last metric we use is the reduction in Value-at-Risk (VaR). As an
RC 0.495 60.138 −1.990 2.923 1.148 39.235 alternative risk measure, VaR has been widely embraced by financial in-
CCC 0.428 65.492 −1.716 3.200 1.079 42.892 stitutions to assess risks (e.g., Basel Committee on Banking Supervision,
DCC 0.364 70.447 −1.457 3.456 0.997 47.157 1996). In essence, VaR estimates the maximum loss a portfolio can incur
BEKK 0.350 71.766 −1.402 3.511 0.975 48.408
over a given time period with a pre-specified probability. Unlike vari-
C-GARCH 0.449 63.815 −1.802 3.110 1.089 41.887
ance which accounts for both gains and losses in a symmetric way and
U.S. thus treats all uncertainty as risk, VaR has asymmetric treatment be-
Unhedged 0.616 – −2.437 – 1.320 –
tween upside risk and downside risk, and considers only losses not
Naïve 0.082 86.702 −0.342 2.095 0.457 65.386
C-OLS 0.090 85.339 −0.364 2.072 0.491 62.747 gains. We want to use it here as a robustness check. To simplify our anal-
C-Cointg 0.073 88.005 −0.305 2.132 0.438 66.786 ysis, we calculate VaR as follows:
TV-OLS 0.081 86.888 −0.329 2.107 0.461 65.041
TV-Cointg 0.075 87.836 −0.308 2.128 0.441 66.534 pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
RC 0.068 88.989 −0.278 2.158 0.421 68.088 VaR ¼ Eðxtþ1 Þ þ zα Var ðxtþ1 Þ ð19Þ
CCC 0.069 88.743 −0.282 2.154 0.428 67.547
DCC 0.070 88.601 −0.286 2.151 0.430 67.365
where zα is the left-tailed critical value at level α of the normal distribu-
BEKK 0.067 89.065 −0.278 2.158 0.418 68.317
C-GARCH 0.072 88.209 −0.295 2.141 0.438 66.775 tion. In practice, α is usually set to be 5% or 1%. Panel C of Table 4 shows
the results. Our focus is on VaR reduction (%), which is the percentage of
Notes: The results are generated by a recursive window scheme. The out-of-sample eval-
uation period is set to be 75% of the full sample. Variance denotes the variance of the port- the VaR reduction that a hedged strategy achieves in comparison with
folio returns (Eq. (16)). Variance reduction measures the relative variance reduction of the the unhedged one. Once again, the relative performance of the models
hedged strategy over the unhedged strategy (Eq. (17)). Utility is the average daily utility is largely in line with that observed using the previous two metrics.
for an investor with a mean-variance utility function (Eq. (18)) and a coefficient of risk
aversion of 4. Utility improvement is the increase in utility by using the hedged strategy
relative to the unhedged strategy. VaR is the Value-at-Risk estimated using (Eq. (19)).
4.2. Statistical significance testing of the relative performance
VaR reduction is the relative VaR reduction of the hedged strategy over the hedged
strategy. An important issue to consider for model selection is data snooping.
According to White (2000), data snooping occurs when a dataset is used
more than once for model selection and inference purposes. In other
portfolio variance. This makes logical sense as it aligns with the ultimate words, using the same dataset repeatedly for testing the performance
risk-reduction goal of hedging. However, it does not imply that portfolio of hedge-ratio estimation models would increase the probability that
return is a completely useless variable. In addition to variance reduc- the results are generated purely due to chance. An inferior model can
tion, some hedgers may wonder how much economic benefit the hedg- be ‘lucky’ and perform better than all other models, and the more
ing can bring. As such, we can use utility gain as a supplementary models that are being compared the higher is the probability that the
measure of hedging performance. Moreover, using it has an added best model has a worse sample performance than some inferior models.
benefit: we can consider the impact of transaction costs (e.g., Alizadeh Given this, we have to evaluate whether the observed superior perfor-
et al., 2008). mance of certain models is indeed significant or occurs by chance. To
696 J. Zhou / Economic Modelling 52 (2016) 690–698

Table 5
MCS test results based on a recursive window scheme for a 75% out-of-sample.

Naïve C-OLS C-Cointg TV-OLS TV-Cointg RC CCC DCC BEKK C-GARCH

Australia 0.012 0.089 0.183 0.183 0.821 0.821 0.089 0.183 1.000 0.052
Europe 0.000 0.029 0.006 0.031 0.031 0.262 0.262 0.262 1.000 0.031
Japan 0.029 0.120 0.029 0.628 0.029 0.029 0.029 0.229 1.000 0.029
U.S. 0.414 0.083 0.513 0.083 0.414 0.857 0.857 0.834 1.000 0.414

Notes: This table presents the p-values of Hansen et al.'s (2011) Model Confidence Set (MCS) test. These results are based on the estimated hedge ratios of Table 4. The relative performance
is defined in Eq. (20). The best models selected with α = 0.10 are highlighted in bold. The block length of bootstrap and the number of bootstrap samples are set equal to 10 and 1000,
respectively.

this end, we employ Hansen et al.'s (2011) Model Confidence Set (MCS) different set of models identified as the best performers. This implies
test.4 that even though BEKK appears to be superior to others in Table 4
The objective of the test is to determine the set of models, M⁎, that when being evaluated using the three metrics, its performance may
consists of the best model(s) from an initial collection of models, M0. not be statistically different than that of some competitors, as shown
MCS is analogous to the confidence interval of a parameter in that it in- here. Finally, despite that a number of models are found to constitute
cludes the best model(s) with a certain probability. Construction of a the cohort of best performers in each market, only two (i.e., BEKK &
MCS involves a sequence of tests for equal predictive ability (EPA). DCC) consistently belong across all markets under study.
This trims the set of candidate models by deleting models that are
found to be significantly inferior. The set of surviving models is the 4.3. Robustness check and implications of findings
MCS, which is guaranteed to contain the best model(s) with a certain
level of confidence. To check the robustness of our results, we try two different out-of-
The MCS procedure starts with a null hypothesis that all candidate sample periods. Tables 6 & 7 report the results for a 50% out-of-
models in M0 have equal performance. Define the relative performance sample based on a recursive window specification. Tables 8 & 9 present
between model i and j as: the results for a 25% out-of-sample but based on a rolling window spec-
ification. If we focus on the big picture without going into the relative
ranking of each model in a certain market, we find that BEKK & DCC
 2  2
^ ΔF t − ΔSt −β
di j;t ¼ ΔSt −β ^ ΔF t : ð20Þ
5 are still consistently among the best performers. This is same as before,
i;t j;t
but now other models seem to share the same trait: C-OLS and TV-OLS
for a 50% out-of-sample while RC for a 25% out-of-sample. However, if
we take together the results from the three different out-of-sample pe-
The null hypothesis then takes the form H 0;M0 : E½di j;t  ¼ 0; ∀i; j ¼ 1; riods, we conclude that only BEKK & DCC are the consistent best
…; M. We use the ‘deviation’ test statistic defined as T D ¼ M−1 ∑i∈M0 t 2i , performers.
pffiffiffi qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
pffiffiffi Given the above findings, what implications do they have? First and
where t i ¼ T di = varð T di Þ represents the standardized relative the easiest, multiple models perform the best in each market and the
performance of model i with respect to the average performance of mix of best performers varies across markets. Such a finding is hardly
the models in Μ0, di ¼ M −1 ∑ j∈M0 di j , and di j ¼ n−1 ∑t¼1 di j;t is a mea-
n
surprising, given the number of markets and of competing methods
sure of relative performance between models i and j. A block bootstrap considered. Second, our results do not suggest that constant hedge-
scheme is used to obtain the distribution under the null. If the null is ratio methods are necessarily inferior. This view is also shared by
rejected, an elimination rule removes the model with the largest ti. Ghosh (1993), Lien et al. (2002) and Moosa (2003), etc. Though the
This process is repeated until non-rejection of the null occurs, thus time-varying methods appear more theoretically appealing, we do
allowing to construct a (1 − α)-confidence set for the best models, Μ*. find that their two constant counterparts (i.e., C-OLS and C-Cointg)
Table 5 presents the p-values of the MCS tests. We follow Hansen often show up among the best performers (see Tables 5, 7 & 9). Third,
et al. (2011) to set α = 0.10. Those models with a p-value greater a more complicated model like C-GARCH does not necessarily yield bet-
than 0.10 are considered not to be outperformed by peers. They consti- ter performance. In some cases (e.g., Australia, Europe, & Japan in
tute the MCS, which contains the best models. They are highlighted in Table 5) C-GARCH clearly loses out to DCC or BEKK while in others
bold. All other models with a p-value equal or lower than 0.10 are infe- (the U.S. in Table 5) its performance is not significantly different than
rior. Given this, we identify the following models for each market that that of DCC or BEKK. Similar conclusions can be drawn using Tables 7
are not statistically outperformed by competitors: C-Cointg, TV-OLS, & 9. Remember our analysis is conducted within an out-of-sample fore-
TV-Cointg, RC, DCC, & BEKK for Australia, RC, CCC, DCC, & BEKK for casting framework. So this reflects a view expressed in the forecasting
Europe, C-OLS, TV-OLS, DCC, & BEKK for Japan, and all models except literature: a complicated model does not always provide better perfor-
C-OLS and TV-OLS for the U.S. It is easy to see that each market has a mance than a parsimonious one (e.g., Weigend and Gershenfeld,
1994; Swanson and White, 1997; Stock and Watson, 1999, and others).
4
Hansen's (2005) Superior Predictive Ability (SPA) can also be used for checking data The underlying argument is that the benefits of using a more complicat-
snooping. Unlike MCS which directly reveals what model(s) are not significantly ed specification may not outweigh the costs caused by its lower degree
outperformed by peers (i.e., the best performers) and what model(s) are outperformed, of parsimony. Note that our conclusion is different than that of Hsu et al.
SPA requires pre-specifying a benchmark model and provides evidence regarding the rel-
(2008), which found that C-GARCH has better hedging performance
ative performance of a particular model. Using SPA leads to qualitatively similar results as
MCS. To preserve space, we do not present the SPA results but they are available upon than those regular GARCH models. Their conclusion is made solely
request. based on variance reduction without testing for the statistical signifi-
5
This function has been first introduced by Lee et al. (2006) to statistically compare cance of the difference between competing methods. Finally, if we
hedging performance. To see how it works, note that the item ðΔSt −β ^ Δ F t Þ2 measures have to report some empirical regularity, we find that only DCC and
i;t
^ Δ F t Þ2 measures the
the portfolio variance based on hedge-ratio model i while ðΔSt −β BEKK are ranked consistently among the best performers across all
j;t

portfolio variance based on model j. Because hedging is to reduce variance, if dij,t, defined four markets when we collectively examine the results based on differ-
as the difference between the two variances, equals 0, this would imply that the two ent out-of-sample periods. This testifies to the popularity of those
models have equal performance. This explains why the null hypothesis is to test whether models (e.g., Baillie and Myers, 1991; Park and Switzer, 1995;
the expected value of dij,t is equal to zero (i.e., E[dij,t] = 0). Kavussanos and Nomikos, 2000; Floros and Vougas, 2006, and others).
J. Zhou / Economic Modelling 52 (2016) 690–698 697

Table 6 Table 8
Hedging effectiveness results based on a recursive window scheme for a 50% out-of- Hedging effectiveness results based on a rolling window scheme for a 25% out-of-sample.
sample.
Panel A Panel B Panel C
Panel A Panel B Panel C
Variance Variance Utility Utility VaR VaR
Variance Variance Utility Utility VaR VaR reduction improvement reduction
reduction improvement reduction (%) (%)
(%) (%)
Australia
Australia Unhedged 0.611 – −2.408 – 1.322 –
Unhedged 0.739 – −2.902 – 1.469 – Naïve 0.335 45.180 −1.336 1.071 0.955 27.720
Naïve 0.222 69.910 −0.901 2.001 0.764 47.962 C-OLS 0.316 48.204 −1.260 1.147 0.931 29.583
C-OLS 0.214 71.033 −0.863 2.039 0.754 48.663 C-Cointg 0.316 48.348 −1.257 1.151 0.929 29.672
C-Cointg 0.213 71.075 −0.862 2.040 0.754 48.684 TV-OLS 0.300 50.900 −1.192 1.215 0.908 31.279
TV-OLS 0.214 71.064 −0.863 2.039 0.754 48.698 TV-Cointg 0.300 50.926 −1.192 1.216 0.908 31.302
TV-Cointg 0.213 71.085 −0.862 2.040 0.754 48.705 RC 0.284 53.492 −1.112 1.296 0.901 31.817
RC 0.199 73.025 −0.799 2.103 0.732 50.146 CCC 0.318 47.891 −1.261 1.147 0.941 28.814
CCC 0.227 69.186 −0.911 1.992 0.785 46.522 DCC 0.289 52.631 −1.136 1.272 0.907 31.386
DCC 0.203 72.647 −0.810 2.092 0.738 49.741 BEKK 0.281 53.992 −1.101 1.306 0.895 32.257
BEKK 0.203 72.497 −0.809 2.093 0.745 49.277 C-GARCH 0.313 48.812 −1.234 1.174 0.937 29.129
C-GARCH 0.218 70.398 −0.877 2.025 0.767 47.763
Europe
Europe Unhedged 0.592 – −2.310 – 1.324 –
Unhedged 1.061 – −4.196 – 1.742 – Naïve 0.226 61.839 −0.917 1.393 0.768 41.985
Naïve 0.405 61.781 −1.631 2.565 1.038 40.417 C-OLS 0.198 66.582 −0.791 1.519 0.732 44.701
C-OLS 0.381 64.009 −1.525 2.671 1.019 41.508 C-Cointg 0.199 66.471 −0.795 1.515 0.732 44.727
C-Cointg 0.380 64.154 −1.520 2.676 1.015 41.700 TV-OLS 0.197 66.651 −0.786 1.524 0.735 44.518
TV-OLS 0.381 64.028 −1.524 2.672 1.018 41.551 TV-Cointg 0.197 66.795 −0.785 1.526 0.731 44.767
TV-Cointg 0.380 64.158 −1.520 2.676 1.015 41.732 RC 0.204 65.531 −0.821 1.489 0.738 44.248
RC 0.391 63.156 −1.571 2.625 1.020 41.419 CCC 0.205 65.363 −0.820 1.490 0.745 43.710
CCC 0.396 62.669 −1.586 2.610 1.033 40.681 DCC 0.205 65.403 −0.823 1.487 0.741 44.045
DCC 0.401 62.154 −1.614 2.582 1.034 40.636 BEKK 0.200 66.186 −0.805 1.505 0.732 44.698
BEKK 0.388 63.382 −1.559 2.636 1.019 41.474 C-GARCH 0.205 65.316 −0.823 1.488 0.744 43.792
C-GARCH 0.408 61.479 −1.638 2.558 1.048 39.835
Japan
Japan Unhedged 0.731 – −2.847 – 1.485 –
Unhedged 1.254 – −4.934 – 1.926 – Naïve 0.166 77.256 −0.681 2.166 0.655 55.918
Naïve 0.451 63.987 −1.807 3.127 1.105 42.622 C-OLS 0.150 79.471 −0.608 2.239 0.630 57.562
C-OLS 0.423 66.237 −1.690 3.244 1.074 44.205 C-Cointg 0.160 78.129 −0.654 2.194 0.644 56.612
C-Cointg 0.450 64.040 −1.803 3.131 1.104 42.676 TV-OLS 0.156 78.669 −0.627 2.221 0.647 56.444
TV-OLS 0.414 66.947 −1.652 3.281 1.065 44.706 TV-Cointg 0.160 78.078 −0.653 2.194 0.647 56.449
TV-Cointg 0.442 64.740 −1.768 3.166 1.095 43.151 RC 0.061 91.679 −0.249 2.599 0.401 73.023
RC 0.569 54.569 −2.284 2.650 1.236 35.833 CCC 0.189 74.176 −0.760 2.087 0.710 52.181
CCC 0.491 60.830 −1.961 2.973 1.157 39.892 DCC 0.106 85.482 −0.436 2.412 0.525 64.651
DCC 0.445 64.522 −1.777 3.157 1.100 42.886 BEKK 0.093 87.271 −0.383 2.464 0.491 66.905
BEKK 0.382 69.494 −1.526 3.408 1.021 46.956 C-GARCH 0.209 71.307 −0.859 1.988 0.734 50.588
C-GARCH 0.521 58.458 −2.088 2.846 1.183 38.543
U.S.
U.S. Unhedged 0.389 – −1.491 – 1.091 –
Unhedged 0.677 – −2.671 – 1.390 – Naïve 0.088 77.319 −0.368 1.122 0.473 56.618
Naïve 0.109 83.856 −0.452 2.218 0.528 62.001 C-OLS 0.069 82.155 −0.280 1.211 0.431 60.492
C-OLS 0.097 85.536 −0.397 2.274 0.509 63.361 C-Cointg 0.072 81.357 −0.297 1.194 0.436 59.983
C-Cointg 0.096 85.887 −0.392 2.279 0.498 64.126 TV-OLS 0.067 82.764 −0.274 1.216 0.419 61.560
TV-OLS 0.097 85.550 −0.398 2.273 0.507 63.490 TV-Cointg 0.071 81.726 −0.291 1.199 0.431 60.486
TV-Cointg 0.096 85.740 −0.395 2.276 0.501 63.909 RC 0.063 83.530 −0.257 1.225 0.414 62.181
RC 0.085 87.347 −0.348 2.323 0.475 65.807 CCC 0.061 84.361 −0.242 1.248 0.407 62.702
CCC 0.085 87.333 −0.350 2.321 0.474 65.846 DCC 0.061 84.164 −0.242 1.248 0.412 62.223
DCC 0.087 87.111 −0.355 2.316 0.479 65.502 BEKK 0.068 82.322 −0.273 1.217 0.433 60.315
BEKK 0.086 87.235 −0.354 2.317 0.475 65.819 C-GARCH 0.062 84.044 −0.247 1.243 0.410 62.392
C-GARCH 0.090 86.637 −0.370 2.301 0.486 65.007
Notes: The results are generated by a rolling window scheme. The out-of-sample evaluation
Notes: The results are generated by a recursive window scheme. The out-of-sample evalua- period is set to be 25% of the full sample. Variance denotes the variance of the portfolio
tion period is set to be 50% of the full sample. Variance denotes the variance of the portfolio returns (Eq. (16)). Variance reduction measures the relative variance reduction of the hedged
returns (Eq. (16)). Variance reduction measures the relative variance reduction of the hedged strategy over the unhedged strategy (Eq. (17)). Utility is the average daily utility for an inves-
strategy over the unhedged strategy (Eq. (17)). Utility is the average daily utility for an inves- tor with a mean-variance utility function (Eq. (18)) and a coefficient of risk aversion of 4.
tor with a mean-variance utility function (Eq. (18)) and a coefficient of risk aversion of 4. Utility improvement is the increase in utility by using the hedged strategy relative to the un-
Utility improvement is the increase in utility by using the hedged strategy relative to the un- hedged strategy. VaR is the Value-at-Risk estimated using (Eq. (19)). VaR reduction is the rel-
hedged strategy. VaR is the Value-at-Risk estimated using (Eq. (19)). VaR reduction is the rel- ative VaR reduction of the hedged strategy over the hedged strategy.
ative VaR reduction of the hedged strategy over the hedged strategy.

Table 7
MCS test results based on a recursive window scheme for a 50% out-of-sample.

Naïve C-OLS C-Cointg TV-OLS TV-Cointg RC CCC DCC BEKK C-GARCH

Australia 0.235 0.271 0.271 0.271 0.271 1.000 0.156 0.440 0.271 0.235
Europe 0.042 0.744 0.945 0.744 1.000 0.744 0.299 0.299 0.744 0.042
Japan 0.049 0.284 0.049 0.420 0.049 0.049 0.049 0.144 1.000 0.049
U.S. 0.262 0.262 0.386 0.243 0.386 1.000 0.991 0.866 0.953 0.386

Notes: This table presents the p-values of Hansen et al.'s (2011) Model Confidence Set (MCS) test. These results are based on the estimated hedge ratios of Table 6. The relative performance
is defined in Eq. (20). The best models selected with α = 0.10 are highlighted in bold. The block length of bootstrap and the number of bootstrap samples are set equal to 10 and 1000,
respectively.
698 J. Zhou / Economic Modelling 52 (2016) 690–698

Table 9
MCS test results based on a rolling window scheme for a 25% out-of-sample.

Naïve C-OLS C-Cointg TV-OLS TV-Cointg RC CCC DCC BEKK C-GARCH

Australia 0.053 0.087 0.096 0.475 0.475 0.483 0.087 0.475 1.000 0.087
Europe 0.040 0.877 0.877 0.877 1.000 0.140 0.140 0.140 0.877 0.140
Japan 0.079 0.299 0.086 0.052 0.086 1.000 0.033 0.499 0.499 0.012
U.S. 0.018 0.018 0.018 0.823 0.772 0.823 1.000 0.823 0.513 0.823

Notes: This table presents the p-values of Hansen et al.'s (2011) Model Confidence Set (MCS) test. These results are based on the estimated hedge ratios of Table 8. The relative performance
is defined in Eq. (20). The best models selected with α = 0.10 are highlighted in bold. The block length of bootstrap and the number of bootstrap samples are set equal to 10 and 1000,
respectively.

However, this does not mean that hedgers will always want to use Gregory, A.W., Hansen, B.E., 1996. Residual-based tests for cointegration in models with
regime shifts. J. Econ. 70, 99–126.
them. For example, in Australia of Table 5 where performances of Hansen, P.R., 2005. A test for superior predictive ability. J. Bus. Econ. Stat. 23, 365–380.
BEKK (or DCC) and C-Cointg are shown not statistically different, Hansen, P.R., Lunde, A., Nason, J.M., 2011. The model confidence set. Econometrica 79 (2),
hedgers may want to use C-Cointg: one, C-Cointg is more computation- 453–497.
Hsu, C.C., Tseng, C.P., Wang, Y.H., 2008. Dynamic hedging with futures: a copula-based
ally friendly than BEKK (or DCC), and two, using C-Cointg avoids the GARCH model. J. Futur. Mark. 28 (11), 1095–1116.
costs of periodically rebalancing the hedged portfolio. Joe, H., 1997. Multivariate Models and Dependence Concepts. Chapman & Hall, London.
Kavussanos, M.G., Nomikos, N.K., 2000. Constant vs. time varying hedge ratios and hedg-
ing efficiency in the BIFFEX market. Transp. Res. E 36, 229–248.
5. Conclusions Kolb, R.W., 2003. Futures, Options, and Swaps. Blackwell Publishing, Malden.
Lee, C.L., Lee, M.-L., 2012. Hedging effectiveness of REIT futures. J. Prop. Invest. Financ. 30
(3), 257–281.
Our study investigates what method is most appropriate to estimate
Lee, H.T., Yoder, J.K., Mittelhammer, R.C., McCluskey, J.J., 2006. A random coefficient
the minimum-variance hedge ratio for the REIT sector. We consider autoregressive Markov regime switching model for dynamic futures hedging.
some most commonly used and relatively new methods. They cover a J. Futur. Mark. 26 (2), 103–129.
Lee, C.L., Stevenson, S., Lee, M.-L., 2014. Futures trading, spot price volatility and market
varying degree of complexity and include both constant and time-
efficiency: evidence from European real estate securities futures. J. Real Estate Financ.
varying versions of hedge ratio. We apply them to all four global REIT Econ. 48, 299–322.
markets which have developed index futures. Our results show that Lien, D., Tse, Y.K., 2002. Some recent developments in futures hedging. J. Econ. Surv. 16
multiple models perform the best in each market and the mix of best (3), 357–396.
Lien, D., Tse, Y.K., Tsui, A.C., 2002. Evaluating the hedging performance of the constant-
performers varies across markets. Furthermore, our results suggest correlation GARCH model. Appl. Financ. Econ. 12, 791–798.
that constant hedge-ratio methods are not necessarily inferior, and Moosa, I.A., 2003. The sensitivity of the optimal hedging ratio to model specification.
that a more complicated GARCH model does not necessarily lead to bet- Financ. Lett. 1, 15–20.
Narayan, Paresh Kumar, Popp, Stephan, 2010. A new unit root test with two structural
ter performance than using a more parsimonious one. Finally, we find breaks in level and slope at unknown time. J. Appl. Stat. 37 (9), 1425–1438.
that only DCC and BEKK rank consistently among the best performers Narayan, Paresh Kumar, Popp, Stephan, 2013. Size and power properties of structural
across all four markets when we examine collectively the results break unit root tests. Appl. Econ. 45 (6), 721–728.
Nelson, R.B., 2006. An Introduction to Copulas. Springer Series in Statistics.
based on different out-of-sample periods. Newell, G., Tan, Y.K., 2004. The development and performance of listed property trust
futures. Pac. Rim Prop. Res. J. 10 (2), 132–145.
Ong, S.E., Ng, K.H., 2009. Developing the real estate derivative market for Singapore:
References issues and challenges. J. Prop. Invest. Financ. 27 (4), 425–432.
Park, T., Switzer, L., 1995. Bivariate GARCH estimation of the optimal hedge ratios for
Alizadeh, A.H., Nomikos, N.K., Pouliasis, P.K., 2008. A Markov regime switching approach
stock index futures: a note. J. Futur. Mark. 15, 61–67.
for hedging energy commodities. J. Bank. Financ. 32, 1970–1983.
Patton, A.J., 2006. Modeling asymmetric exchange rate dependence. Int. Econ. Rev. 47,
Baillie, R.T., Myers, R.J., 1991. Bivariate GARCH estimation of the optimal commodity
527–556.
futures hedge. J. Appl. Econ. 6, 109–124.
Patton, A.J., 2009. Copula-based models for financial time series. In: Anderson, T.G. (Ed.),
Basel Committee on Banking Supervision, 1996. Overview of the Amendment to the
Handbook of Financial Time Series. Springer, pp. 767–781.
Capital Accord to Incorporate Market Risks available at http://www.bis.org.
Phan, Dinh Hoang Bach, Sharma, Susan Sunila, Narayan, Paresh Kumar, 2015. Stock return
Bera, A.K., Garcia, P., Roh, J.S., 1997. Estimation of time-varying hedging ratios for corn and
forecasting: some new evidence. Int. Rev. Financ. Anal. 40, 38–51.
soybeans: BGARCH and Random Coefficient approaches. Indian J. Stat. B 59, 346–368.
Popp, S., 2008. New innovational outlier unit root test with a break at an unknown time.
Bollerslev, T., 1990. Modeling the coherence in short-run nominal exchange rates: a
J. Stat. Comput. Simul. 78, 1143–1159.
multivariate generalized ARCH model. Rev. Econ. Stat. 72, 498–505.
Rogoff, K., Stavrakeva, V., 2008. The continuing puzzle of short horizon exchange rate
Bos, T., Newbold, P., 1984. An empirical investigation of the possibility of stochastic
forecasting. NBER Working Paper No. w14071.
systematic risk in the market model. J. Bus. 57 (1), 35–41.
Shi, J., Xu, T., 2013. Price and volatility dynamics between securitized real estate spot and
Chen, S.-S., Lee, C.-F., Shrestha, K., 2003. Futures hedge ratios: a review. Q. Rev. Econ.
futures markets. Econ. Model. 35, 582–592.
Finance 43, 433–465.
Sklar, A., 1959. Fonctions de Réparitition à n dimensions et Leurs Marges. Publications de
Durbin, J., Koopman, S.J., 2001. Time Series Analysis by State Space Methods. Oxford
l'Institut de Statistique de l'Université de Paris 8, pp. 229–231.
University Press.
Stock, J.H., Watson, M.W., 1999. A comparison of linear and nonlinear univariate models
Ederington, L.H., 1979. The hedging performance of the new futures markets. J. Financ. 34,
for forecasting macroeconomic time series. In: Engle, R.F., White, H. (Eds.),
157–170.
Cointegration, Causality, and Forecasting: A Festschrift in Honor of Clive W. J.
Engle, R.F., 2002. Dynamic conditional correlation: a simple class of multivariate GARCH
Granger. Oxford University Press, Oxford.
models. J. Bus. Econ. Stat. 20, 339–350.
Swanson, N.R., White, H., 1997. A model selection approach to real-time macroeconomic
Engle, R.F., Kroner, F.K., 1995. Multivariate simultaneous generalized ARCH. Econ. Theory
forecasting using linear models and artificial neural networks. Rev. Econ. Stat. 79,
11, 122–150.
540–550.
Figlewski, S., 1984. Hedging performance and basis risk in stock index futures. J. Financ.
Tsay, R.S., 2005. Analysis of Financial Time Series. Second Edition. Wiley.
39, 657–669.
Weigend, A.S., Gershenfeld, N.A., 1994. Time Series Prediction: Forecasting the Future and
Floros, C., Vougas, D.V., 2006. Hedging effectiveness in Greek Stock index futures market
Understanding the Past. Addison-Wesley, Boston.
1999–2001. Int. Res. J. Financ. Econ. 5, 7–18.
White, H., 2000. A reality check for data snooping. Econometrica 68, 1097–1126.
Ghosh, A., 1993. Cointegration and error correction models: intertemporal causality be-
tween index and futures prices. J. Futur. Mark. 13, 193–198.

You might also like