Dynamic Hedging Performance of the CSI 300 Index Futures - The ...

5 downloads 0 Views 401KB Size Report
Dec 5, 2014 - performance of index futures in China's financial markets, but none of .... low-frequency models for the S&P 500 and the FTSE 100 indices as ...
Dynamic Hedging Performance of the CSI 300 Index Futures - The Realized Minimum-Variance Hedge Ratio Approach Hui Qu School of Management and Engineering Nanjing University, China Tel: 86-13951604813; Email: [email protected]

Tianyang Wang (Corresponding Author) Finance and Real Estate Department Colorado State University, USA Tel: 1-9704912381; Email: [email protected]

Yi Zhang School of Management and Engineering Nanjing University, China Tel: 86-13770680103; Email: [email protected]

Pengfei Sun School of Management and Engineering Nanjing University, China Tel: 86-15298386501; Email: [email protected]

1

Dynamic Hedging Performance of the CSI 300 Index Futures

2

- The Realized Minimum-Variance Hedge Ratio Approach

3 4

Abstract: This paper comprehensively investigates the dynamic hedging performance

5

of China’s CSI 300 index futures by using the realized minimum-variance hedge ratio

6

(RMVHR) as an efficient way to utilize the high-frequency intraday information. We

7

thoroughly examine a number of RMVHR-based time-series models for CSI 300 index

8

futures, and evaluate the out-of-sample dynamic hedging performance in comparison

9

to the conventional hedging models using both lower frequency and high frequency

10

data. Our results show that the dynamic hedging performance of the RMVHR-based

11

methods robustly dominates that of the conventional methods in terms of major

12

performance measures including the hedge ratio, the hedging effectiveness, the

13

portfolio return and the Sharp ratio in the out-of-sample forecast period. Furthermore,

14

the superiority of the RMVHR-based methods is consistent during different volatility

15

regimes of China’s financial markets, including the market turbulence in 2015.

16

Keywords: Realized Minimum-Variance Hedge Ratio; Dynamic Hedging Performance,

17

High-Frequency Data; Hedging Effectiveness; Volatility Regime

18 19

1

Introduction

20

The launch of the CSI 300 equity index futures on April 16, 2010 marked a

21

milestone development in the evolution of China’s financial market. For the first time,

22

China’s financial market provides investors with an essential tool to hedge the

23

systematic risk of holding the market, proxied by the underlying CSI 300 equity index,

24

a free-float weighted index comprises 300 of the largest and actively traded A-share

25

stocks on the Shanghai and Shenzhen Stock exchanges.

26

While its inception was widely hailed as an effective hedging tool and even a

27

stabilizing force in China’s financial markets among investors and regulators 1 , the

28

effectiveness of the hedging strategy using the CSI 300 index futures contracts in the

29

dynamic settings has been a subject of very limited academic research. Despite the

30

growing importance of stock index futures and almost eight years since its introduction

31

in China, the existing literature on its hedging effectiveness mainly focuses on the

32

conventional constant hedge ratios with low frequencies of data (mostly daily) over

33

relatively short hedging horizons (c.f. Zhang and Shen 2011; Wen, Wei, and Huang

34

2011; He and Yang 2012; Hou and Li 2013; Pan and Sun 2014), and hence is subject to

35

important weaknesses. First, many criticisms have been suggested against the

36

conventional hedge ratios methods used in the early literature because they can’t fully

37

incorporate the dynamic hedging relation and the past available information (e.g.,

38

Baillie and Myers, 1991; Yang and Allen, 2004; Park and Jei, 2010). Second, empirical

39

study has suggested valuable informational content of high-frequency intraday data in

40

estimating hedge ratios over the low frequency data (e.g. Lai and Sheu 2010; Han and

41

Ren 2012; Markopoulou et al. 2016; Miao et al. 2017). In addition, the China’s stock

42

market turbulence in 2015 provides an interesting window to examine the hedging

43

performance of index futures in China’s financial markets, but none of the earlier

44

studies evaluates the hedging performance of the CSI 300 index futures during the

45

turbulent market fluctuations. This is especially important as the China Securities

46

Regulatory Commission (CSRC) openly blamed the stock market collapse on

47

“malicious short-selling” of index futures as “weapons of mass destruction” by

On December 5, 2014, Xiao Gang, chairman of the China Securities Regulatory Commission (CSRC) remarked, stock index futures are “sophisticated risk management tools for improving the stock market operation mechanism, providing hedging instruments, improving the investment product market system and promoting stable development of great significance.” 1

1

48

speculators and questioning its conventional role as a hedging instrument. Therefore, a

49

clear picture of the dynamic hedging performance of the CSI 300 index futures is

50

essential for investors and calls for close attention from the academics.

51

The new technical development in hedging ratio measures, the increasing

52

availability of high-frequency intraday data, and the tumultuous market environment in

53

China’s recent financial market motivate us to assess and examine the hedging

54

performance of the new proposed realized minimum-variance hedge ratio (RMVHR)

55

(Markopouloua, et al. 2016), constructed from high frequency data (five-minute

56

frequency) of China’s CSI 300 index and index futures. Our study shows that the

57

dynamic hedging performance of the RMVHR-based methods dominates that of the

58

conventional methods in terms of major performance measures including the hedge

59

ratio, the hedging effectiveness, the portfolio return and the Sharp ratio in the out-of-

60

sample forecast period. Furthermore, the superiority of the RMVHR-based methods is

61

consistent during different volatility regimes of China’s financial markets, including

62

China’s market turbulence in 2015.

63

This research contributes to the existing literature in the following four dimensions.

64

First, to our best knowledge, this is the first study to comprehensively investigate

65

the dynamic hedging performance of CSI 300 index futures by applying and extending

66

the new proposed realized minimum-variance hedge ratio. In contrast to earlier studies

67

that use relatively low frequency and short examination windows, we use the intraday

68

five-minute data of CSI 300 index and index futures, and employ a variety of time-

69

series models to forecast the RMVHR ratio from a much longer sample period. Our

70

study shows that our proposed method is more efficient than conventional models in

71

terms of major performance measures including the hedge ratio, the hedging

72

effectiveness, the portfolio return and the Sharp ratio criteria. 2

73

Second, we propose a method to directly measure the marginal benefits of using

74

RMVHR in China’s financial markets and show that constructing the RMVHR and

75

directly forecasting it is a more efficient way to utilize the high-frequency intraday

76

information. More specifically, in addition to the conventional low-frequency models

77

in the comparison group, we also assess the hedging performance of the dynamic

78

conditional correlation (DCC)-RV model (Lai and Sheu 2010) and the Vector

79

Heterogeneous Autoregressive (VHAR) model (Busch et al., 2011) that also utilize

80

high-frequency data. Because the VHAR model of the realized covariance (RCov)

81

matrix and the heterogeneous autoregressive (HAR) model (Corsi, 2009) of RMVHR

82

utilize exactly the same information set (intraday five-minute returns of spot and futures)

83

and have similar structures, the comparison provides direct measure of the marginal

84

benefit of RMVHR and illustrates its superiority in utilizing the high-frequency

85

intraday information.

86

Third, while none of the earlier studies examine the crisis period in 2015, we

87

carefully test the robustness of our results to different market conditions including

88

China’s turbulent market fluctuations in the out-of-sample forecast period. Using the

89

nonparametric change point model (Ross et al. 2011), different volatility regimes of the

90

underlying CSI 300 index were detected in the forecast period. We show that the

91

superiority of the hedging performance of CSI 300 index futures using RMVHR is

92

robust to different volatility regimes.

93

Last but not the least, observing that the residuals for the HAR model of RMVHR

94

have significant ARCH effects and non-zero skewness, we further extend the HAR

95

model by modeling the variance of the dynamic hedge ratio with the GARCH

96

specification and skewed-t innovations. We show that the resulting HAR-GARCH 3

97

model of RMVHR has lower hedge ratio, competitive hedging effectiveness, much

98

higher portfolio return and Sharp ratio than the basic HAR model, and its superiority is

99

robust to different market conditions.

100

The remainder of the paper is organized as follows. Section 2 provides a brief

101

overview of the literature and draws comparison with our study. Section 3 presents

102

methodology of the RMVHR models and its comparison models. Section 4 explains

103

the data and the results are discussed in Section 5. This is followed by a discussion of

104

robustness tests in Section 6. Section 7 concludes the paper.

105

2. Literature Review

106

The effectiveness of a hedging strategy using futures contracts depends heavily on

107

the accuracy of the optimal hedge ratio estimation. The view of determining hedge ratio

108

that prevails today originates from Johnson (1960) and Stein (1961), who apply the

109

Markowitz modern portfolio theory and propose the minimum-variance method with

110

the objective to minimize portfolio risk. Following them, Ederington (1979) proposes

111

to calculate constant hedge ratio using ordinary least squares (OLS) model. In addition,

112

Ghosh (1993) proposes the error correction (ECM) model for determining hedge ratio

113

based on the cointegration theory, which considers both the long-term equilibrium and

114

the short-term dynamics between spot and futures. However, both models are static

115

hedging methods since they assume constant variance and covariance between spot and

116

futures returns over time and do not consider the time-varying (co)variance of spot and

117

futures.

118

With

the

development

of

the

generalized

autoregressive

conditional

119

heteroscedastic (GARCH) models, dynamic hedging based on the bivariate GARCH

120

models have gained rising attentions in the literature. The bivariate GARCH-ECM 4

121

model (Kroner and Sultan 1993) considers the cointegration relationship and

122

characterizes the time-varying covariance of spot and futures. It has been shown to have

123

better dynamic hedging performance than the static hedging methods, and thus gained

124

wide applications. Alternative common choices of the bivariate GARCH model include

125

the BEKK-GARCH model (Engle and Kroner 1995), the constant conditional

126

correlation (CCC)-GARCH model (Bollerslev 1990), the DCC-GARCH model (Engle

127

2002) and copula-GARCH model (Hsu et al. 2008; Lai et al. 2009). However, these

128

GARCH-class models are likely to overestimate the persistence in volatility since

129

relevant sudden changes in variance are ignored (Wei et al. 2011). In addition, the early

130

studies mainly use relatively low frequency (daily in most cases) data to latently

131

characterize the time-varying covariance of spot and futures, therefore they cannot

132

capture the intraday variation of prices and are slow in catching up the covariance

133

changes.

134

The harnessing of high-frequency information and the new development in

135

financial econometrics have enabled significant progress in direct measuring and

136

modeling of covariance. The realized volatility (RV) calculated as the sum of squared

137

intraday returns provides an unbiased estimator of the quadratic variation (Andersen

138

and Bollerslev, 1998). As a natural extension of RV into the multivariate case, the

139

realized covariance (RCov) matrix calculated as the sum of the cross products of high-

140

frequency intraday return vectors provides an unbiased estimator of the quadratic

141

covariation (Barndorff-Nielsen and Shephard, 2004). Unfortunately, the calculation of

142

the RCov matrix suffers from the problems of market microstructure noise and

143

nonsynchronous trading. While more complicated estimators have been proposed (e.g.,

144

the multivariate realized kernel (Barndorff-Nielsen et al., 2011), the incorporation of

145

investor sentiment and weather effects (Yang, Jhang and Chang 2016), and the two5

146

time scale covariance (Zhang, 2011), their computation complexities impede wide

147

applications in practice.

148

As a more practical alternative, the easily implemented sparse sampling method

149

using high frequencies of data has been employed in empirical applications. For

150

instance, Lai and Sheu (2010) proposed the DCC-RV model using 15-minutes

151

frequency of data, which encompasses the realized volatility (correlation) in the

152

conditional variance (correlation) functions for spot and futures and shows substantial

153

improvement in hedging performance for the S&P 500 index futures. Han and Ren

154

(2012) proposed the vector heterogeneous autoregressive (VHAR) type of model

155

(Busch et al., 2011) which directly models the realized covariance matrix, and achieves

156

better hedging performance compared with the bivariate GARCH-ECM type models.

157

Most recently, Markopoulou et al. (2016) proposed a new concept of the realized

158

minimum-variance hedge ratio (RMVHR) employing high-frequency data. RMVHR is

159

defined as the ratio of the realized covariance between futures and spot returns divided

160

by the realized volatility of futures. In contrast, all former mentioned dynamic hedging

161

methods determine the dynamic hedge ratio as the ratio of the forecasted covariance of

162

futures and spot returns divided by the forecasted futures variance. Markopoulou et al.

163

(2016) show that directly modeling and forecasting the RMVHR can result in

164

substantial hedging performance improvements when compared with conventional

165

low-frequency models for the S&P 500 and the FTSE 100 indices as well as the

166

EUR/USD and the GBP/USD foreign exchange rates. However, it is unclear if the

167

RMVHR would also be beneficial in China’s market given several distinguishing

168

elements in the Chinese stock market, including trading restrictions such as price-limit

169

rules, margin trading, short-selling restrictions and T+1 trading constraints, closed

170

nature of the market to foreign shareholders, and a large portion of retail investors in 6

171

contrast to developed markets.

172

It is important to highlight how our study differs from the previous research on

173

hedging performance of CSI 300 index futures in China’s market. In China’s market,

174

preliminary research has been conducted to assess the hedging performance of CSI 300

175

index futures using constant hedge ratios (cf. Hou and Li 2013), relatively low

176

frequency (mostly daily) (c.f. Zhang and Shen 2011; Wen, Wei, and Huang 2011) and

177

short examination windows surrounding the introduction of equity futures in China in

178

2010 (c.f. Wei et al. 2011; Hou and Li 2013; Pan and Sun 2014). A notable downside

179

of these prior studies is their reliance on lower frequency data, relatively short sample

180

periods and examination of periods that does not fully reflect the broader evolution of

181

the equity index and futures market. More importantly, none of the earlier studies

182

examine the crisis period in 2015, when the Chinese regulatory authority officially

183

undertook a series of regulatory interventions and launched a formal investigation into

184

the role played by the futures market in exacerbating the crisis in response to the stock

185

market crash in June 2015, which as a result raises the question of the conventional

186

hedging role of the index futures in China’s market.

187

We believe that our study offers the most comprehensive investigation of the

188

dynamic hedging performance of CSI 300 index futures in the Chinese stock market.

189

Our analysis benefits from a longer sample period that spans January 2012 through

190

December 2017, which captures the overall evolution of the market as evinced by its

191

steep rise and dramatic decline. We applied and further extended the newly proposed

192

RMVHR approach to examine the dynamic hedging performance of CSI 300 index

193

futures throughout the long horizon of up and down in Chinese stock market.

194

Additionally, our study employs intraday data measured at 5-minutes intervals thus

195

providing an added level of granularity as well as robustness to the analysis. 7

196

3. Methodology

197

3.1 Realized Measures

198

Let the discretely sampled Δ-period log return be denoted by rt+j·Δ,Δ = lnpt+j·Δ -

199

lnpt+(j-1)·Δ, j = 1, 2, …, M, t = 0, 1, 2, …, where pt+j·Δ is the high-frequency price observed

200

at time j·Δ within day t+1 and M = 1/Δ is the number of sampling intervals per day. The

201

daily realized volatility is defined by the summation of the squared intraday returns as

202

RVt ( Δ) ≡ ∑ j =1 rt 2-1+ j ⋅ Δ, Δ (Andersen and Bollerslev, 1998), which converges uniformly

203

in probability to the quadratic variation as Δ→0.

1/ Δ

204

Let rt+j·Δ,Δ = [rSt+j·Δ,Δ, rFt+j·Δ,Δ]’ be the column vector of returns, where rSt+j·Δ,Δ is the

205

day (t+1) Δ-period log return of the CSI 300 index and rFt+j·Δ,Δ is the day (t+1) Δ-period

206

log return of the CSI 300 index futures. The daily realized covariance matrix is defined

207

by the summation of the cross products of intraday return vectors as

208

RCov tS , F ( Δ) ≡ ∑ j =1 rt -1+ j ⋅ Δ, Δ ⋅ rt′-1+ j ⋅ Δ, Δ (Barndorff-Nielsen and Shephard, 2004), which

209

converges uniformly in probability to the quadratic covariation as Δ→0.

210

1/ Δ

The minimum-variance hedge ratio of day t can be calculated as H tS Cov( RtS , RtF ) S ,F , where RtS and RtF are the day t log returns of the spot = ρ × t F F Var ( Rt ) Ht

211

= HR t

212

and the futures, respectively. ρtS,F is the day t correlation between the spot and the

213

futures returns, and HSt and HFt are the day t variances of the spot and the futures,

214

respectively. According to this, the day t realized minimum-variance hedge ratio

215

(RMVHR) is defined as RMVHRt ( Δ) ≡

216

F RCovtS , F ( Δ) is the sub-diagonal element of RCov tS , F ( Δ) , and RVt ( Δ) is the day t

217

realized variance of the futures. For notational simplicity, we omit the notation ( Δ) in

RCovtS , F ( Δ) (Markopulou et al., 2016), where RVt F ( Δ)

8

218

the realized measures when presenting the forecasting models.

219

3.2 Forecasting models

220

As for the RMVHR forecasting models, we consider the following time-series

221

models:

222

1) The ARMA model: = . RMVHRt = c + ∑ i 1ϕ i RMVHRt −i + ∑ j 1ϑ j ε t − j + ε t =

223

2) The ARMA-GARCH model:

p

q

RMVHR = c + ∑ ϕ RMVHR + ∑ ϑ j ε t − j + ε t , p

q

t i t −i = i 1= j 1

224

ε t =σ t et , 2 2 2 σ= ω + ∑ k 1α t = k ε t − k + ∑ l 1 β lσ t − l . = m

n

225

cst + ϕ st RMVHRt −1 + ε t , 3) The Regime-switching (RS) model: RMVHRt =

226

where st is the state variable that takes the values 1 and 2. The state transitions are given

227

pi , j P= st −1 i ) , i, j = 1,2. by a Markov chain with transition probabilities = ( st j |=

228

(

)

(

)

4) The ARFIMA model: 1 − ∑ ϕ L (1 − L ) ( RMVHR − µ ) = 1 + ∑ ϑ j Lj ε t ,

p q d i i t =i 1 = j 1

229

where d is the differencing order and L is the lag operator.

230

5) The HAR model:, RMVHRt = α 0 + α d RMVHRt −1 + α w RMVHRt(−w1) + α m RMVHRt(−m1) + ε t

231

where

232

weekly and monthly RMVHRs.

233

6) The HAR-GARCH model:

RMVHRt(−w1) =

1 5 1 22 RMVHRt −i , RMVHRt(−m1) = ∑ i =1 RMVHRt −i are the past ∑ i =1 5 22

9

234

235

236

RMVHRt = α 0 + α d RMVHRt −1 + α w RMVHRt(−w1) + α m RMVHRt(−m1) + ε t ,

ε t =σ t et , 2 2 2 σ= ω + ∑ k 1α t = k ε t − k + ∑ l 1 β lσ t − l , = m

n

where et follows skewed-t distribution.

237

As for the conventional hedging approaches, we include the static OLS and ECM

238

models, the dynamic DCC model, DCC-RV model and the VHAR model. The former

239

three models completely rely on the daily log returns of the spot (RtS) and the futures

240

(RtF). The DCC-RV model incorporates high-frequency based realized covariance

241

matrix (volatilities and correlation) in the DCC framework; while the VHAR model

242

directly models the high-frequency based realized covariance matrix.

243

α β RtF + ε t . 7) The OLS model: RtS =+

244

α β RtF + γ RtS-1 − θ RtF−1 + ε t , 8) The ECM model: RtS =+

245

where

246

equilibrium between the spot and the futures.

247

)

(

(R

S t -1

− θ RtF−1 ) is the error correction term that characterizes the long-term

The OLS model and the ECM model are all static models, and the estimated

248

parameter β is the (constant) hedge ratio.

249

9) The DCC model: RtS = µ S + γ S ( RtS-1 − θ RtF−1 ) + ε tS ,

250

RtF = µ F + γ F ( RtS-1 − θ RtF−1 ) + ε tF ,  ε tS   F  | ψt −1 ~ N (0, H t ),  εt  10

251

where ψt-1 is the information set up to day (t-1) and Ht is the conditional covariance

252

matrix modeled as: S H tS , F   H t =  H tF   0  S S S S 2 S β 0 + β1 ε t −1 +β 2 H tS−1 , H= t

0   1 × , S F H tF   ρt

 HS H t =  St, F  Ht 253

× 1   0 

0   = Dt R t Dt , H tF 

(*)

F β 0F + β1F ε tF−12 +β 2F H tF−1 , H= t 1

S ρtS , F   H t

(*) 1

R t = diag ( Qt ) 2 Qt diag ( Qt ) 2 , −



Qt = (1 − α − β )Q + α z t −1z t −1′ + β Qt −1 ,

εS

(*)

H tS   is the standardized residual vector, and Q is the unconditional H tF 

254

where z t = 

255

correlation matrix of the spot and the futures returns. α and β are nonnegative scalars

256

with α + β ≤ 1 .

257

10) The DCC-RV model has similar formulation compared to the DCC model, with

258

t

ε F  t

modifications in the three equations of 9) that are marked with (*): S H= β 0S + β1S RVt −S1 +β 2S H tS−1 , t

259

(*)

F H= β 0F + β1F RVt −F1 +β 2F H tF−1 , t

(*)

Qt = (1 − α − β )Q + α RCorrtS−1, F + β Qt −1 , 260

where RCorrtS , F is the realized correlation matrix whose sub-diagonal element is S ,F

261

calculated as RCorrt

262

11) The VHAR model:

263 264

(*)

=

RCovtS , F RVt S ⋅ RVt F

.

The matrix logarithm transformation method is adopted to guarantee the positive definiteness

of

the

forecasted

covariance 11

matrix.

Specifically,

define

265

= X t vech( = At ) A t = logm(RCOVtS , F ) and define

266

is constructed as:

267

 X tS   S ,F  Xt  =  X tF   

 α S   β11( d )  S ,F   (d )  α  +  β 21  α F   β31( d )   

 β11( w)  +  β 21( w)  β31( w) 

β12( w) β 22( w) β32( w)

268

where

269

X tF−1( m ) =

270

X tS−1( w) =

β12( d ) β 22( d ) β32( d )

1 22 X tF−i ∑ i =1 22

X tS−,1F ( w) =

S t

, X tS , F , X tF ) . The VHAR model '

β13( d )   X tS−1    β 23( d )   X tS−,1F  β33( d )   X tF−1 

β13( w)   X tS−1( w)   β11( m )    β 23( w)   X tS−,1F ( w)  +  β 21( m ) β33( w)   X tF−1( w)   β31( m )

1 5 S ∑ X t −i 5 i =1

(X

,

β12( m ) β 22( m ) β32( m )

X tF−1( w) =

β13( m )   X tS−1( m )   ε tS      β 23( m )   X tS−,1F ( m )  +  ε tS , F  , β33( m )   X tF−1( m )   ε tF 

1 5 X tF−i ∑ i =1 5

,

X tS−1( m ) =

22 1 X tS−i , ∑ i =1 22

1 5 S ,F 1 22 X t −i , X tS−,1F ( m ) = ∑ i =1 X tS−,iF . ∑ i =1 22 5

The inverse of the vech() function and the matrix exponential transformation is then

271

applied to get the prediction of the covariance matrix.

272

4.

Data Description

273

Our empirical data consist of five-minute (1/Δ = 48) prices of China’s CSI 300

274

index and index futures from January 4, 2012 to December 29, 2017, covering a total of

275

1456 trading days. 2 We chose the five-minute sparse sampling approach following the

276

majority of previous studies as it provides a good trade-off between accuracy and market

277

microstructure noise (nonsynchronous trading). The trading time of the CSI 300 index

278

futures used to be 9:15am – 11:30am for morning session and 13:00pm – 15:15pm for

279

afternoon session. Since January 1, 2016, China Financial Futures Exchange has

2 There are 1458 trading days from January 4, 2012 to December 29, 2017. However, trading on January 4, 2016 and January 7, 2016 closed much earlier, due to the circuit breaker mechanism being triggered. Thus these two days are deleted from our sample. 12

280

adjusted the opening and closing times for the CSI 300 index futures to 9:30am and

281

15:00pm, respectively, to match those of the CSI 300 index. Thus in this empirical

282

research, we use the five-minutes prices between 9:30am – 11:30am and 13:00pm –

283

15:00pm for both the CSI 300 index and the CSI 300 index futures, deleting all price

284

records in the non-overlapping periods for consistency and comparison purpose.

285

[Insert Figure 1 Here]

286

Figure 1 displays the time series plots of the log daily prices for the CSI 300 index

287

and the CSI 300 index futures in the whole sample period. It shows that the log daily

288

prices of the CSI 300 index futures are very close to those of the CSI 300 index in most

289

of the trading days, and that the Chinese stock market has gone through relative tranquil

290

and extremely volatile periods during our sample period. This observation inspires us to

291

test the robustness of our results to different market conditions, which will be explained

292

later.

293

[Insert Figure 2 Here]

294

Figure 2 displays the time series plots of the realized volatilities for the CSI 300

295

index (RVtS) and the CSI 300 index futures (RVtF) as well as the realized covariance

296

between the spot and the futures (RCovtS,F) in the whole sample period. It shows that the

297

realized volatility of the CSI 300 index futures has a similar pattern as that of the CSI

298

300 index, although it is more volatile. Both the realized volatility series and the realized

299

covariance series are relatively tranquil during the period from January 4, 2012 to the

300

end of 2014, but are very turbulent around the year of 2015. Such pattern necessitates

301

our robustness check in different volatility regimes.

302

[Insert Table 1 Here] 13

303

Table 1 reports descriptive statistics for the realized volatilities (RVtS and RVtF), the

304

realized covariance (RCovtS,F), and the realized minimum-variance hedge ratio

305

(RMVHRt) of the CSI 300 index and index futures over the entire sample period. We can

306

see that the realized volatility of the CSI 300 index futures has higher standard deviation

307

than that of the CSI 300 index, indicating that the CSI 300 index futures is more volatile.

308

The ADF and PP test statistics show that these four realized measures are all stationary,

309

and thus can all be directly modeled. The Ljung-Box test statistics show that these four

310

realized measures all exhibit up to 20th order serial correlation, and thus the long-

311

memory models may be appropriate choices to model the RMVHR and the RCov matrix.

312

[Insert Table 2 Here]

313

Table 2 reports the diagnostic statistics for the regression residuals of the HAR

314

model of RMVHR in the whole sample period from January 4, 2012 to December 29,

315

2017. The ARCH-LM tests indicate strong autoregressive conditional heteroscedasticity

316

effects, which calls for GARCH model for the residual variance. The residuals have

317

positive skewness and kurtosis higher than 3, and are significantly non-normal as

318

confirmed by the Jarque-Bera normality test statistics. Thus, we assume skewed-t

319

innovations.

320

5.

Hedging Performance Comparison

321

We set the period from January 2, 2014 to December 29, 2017 (975 trading days) as

322

the out-of-sample forecast period, and perform one-step-ahead rolling window forecast.

323

That is, we use the period from January 4, 2012 to December 31, 2013 (2 years, 481

324

trading days) as the first estimation window, to make forecasts for January 2, 2014. The

325

estimation window is then rolled forward, and we use the period from January 5, 2012 14

326

to January 2, 2014 as the second estimation window, to make forecasts for January 3,

327

2014. The estimation window keeps rolling forward, until we have made forecasts for

328

all the 975 out-of-sample trading days.

329 330

Based on these forecasts, we perform dynamic hedging of the CSI 300 index futures, and calculate the following four hedging performance indicators:

331

(1) Hedge Ratio (HR): HR = E ( HRt ) , where E() means taking expectation. For the

332

RMVHR-based models, HRt is the predicted day t hedge ratio; for the OLS model

333

and the ECM model, HRt is the predicted coefficient βˆ ; for the DCC model

334

H tS Cov( RtS , RtF ) S ,F and the DCC-RV model, HRt is calculated as HR ; = = × ρ t t F F Var ( Rt )

Ht

RCovtS , F ; for the NAÏVE RVt F

335

for the VHAR model, HRt is calculated as HRt =

336

method , HRt is always equal to 1.

337

The hedge ratio indicates the cost and leverage of the hedging, and thus lower HR

338

is preferred.

339

(2) Hedging Effectiveness (HE): HE = E ( HEt ) , HEt = 1 −

2 σ HP 2 ,t , where σ UP ,t is 2 σ UP ,t

340

the day t variance of the un-hedged portfolio, and is calculated as the realized

341

2 variance of the CSI 300 index ( RVt S ), σ HP ,t is the day t variance of the hedged

342

portfolio, and is calculated using the realized variances of the CSI 300 index and

343

index futures ( RVt F ), and the realized covariance of the spot and the futures (

344

2 RCovtS , F ): σ HP RVt S − 2 βˆt RCovtS ,F + βˆt 2 RVt F , where βˆt is the forecasted ,t =

345

minimum-variance hedge ratio for day t. 15

346

The hedging effectiveness indicator assesses the variance reduction rate of the

347

hedged portfolio. Thus, higher HE is preferred since it means that the portfolio

348

risk has been largely reduced.

349 350 351

= RtS − βˆt RtF , where RtHP is the (3) Portfolio Return (PR): PR = E ( RtHP ) , RtHP day t log return of the hedged portfolio. Naturally, higher PR is preferred. (4) Sharpe Ratio (SR): SR =

PR − R f

σ HP

, where σ HP is the standard deviation of the

352

hedged portfolio returns, R f is the risk-free return which we assume 0 here. This

353

indicator takes into consideration the returns and risks at the same time. Similar

354

to the evaluation of PR, higher SR is preferred.

355

[Insert Table 3 Here]

356

Table 3 reports the hedging performance of all the models in the out-of-sample

357

forecast period from January 2, 2014 to December 29, 2017. It is divided into two

358

panels. Panel I displays results for those models that directly model the RMVHR. Panel

359

II displays results for those models that model the daily returns (covariance matrix).

360

The performance of the NAÏVE method that uses a hedge ratio equal to 1 is also

361

reported in Panel II. In each panel, the hedging performance indicators are listed in the

362

first column, while the models are specified in the second row. For each specified

363

hedging performance indicator, the results of all the twelve methods are compared, and

364

the best result is in bold. For example, the number 62.7216% in Panel I is in bold,

365

because the RS using RMVHR model has the lowest HR indicator value of 62.7216%

366

among all these twelve methods.

367

From Table 3 we can see that when the hedge ratio indicator (HR) is considered, 16

368

the magnitude of the numeric numbers in Panel I using RMVHR are all much smaller

369

than the corresponding numbers in Panel II using alternative methods. When the

370

hedging effectiveness indicator (HE) is considered, the numeric numbers in Panel I are

371

mostly larger than those numbers in Panel II with only one exception, i.e., the RS model

372

in Panel I has slightly lower HE than the VHAR model in panel II. When the portfolio

373

return indicator (PR) or the Sharp ratio indicator (SR) is considered, the numeric

374

numbers in Panel I are all much larger than the corresponding numbers in Panel II.

375

Therefore, our empirical test shows that the dynamic hedging performance of CSI 300

376

index futures using RMVHR dominates that of the conventional methods in terms of

377

all these four hedging performance indicators in the out-of-sample forecast period.

378

These six RMVHR models all have quite close hedging effectiveness, among which the

379

RS model has the lowest hedge ratio, the highest portfolio return and the highest Sharp

380

ratio, and thus is superior in terms of these three economic criteria.

381

Additionally, in Panel II, the DCC-RV model has lower HR, higher HE, higher PR

382

and higher SR than the DCC model, which means that incorporating the information in

383

the realized covariance matrix can effectively improve the dynamic hedging

384

performance. The VHAR model has the lowest HR, the highest HE, the highest PR and

385

the highest SR in Panel II, which indicates that directly modeling the realized

386

covariance matrix can better utilize the intraday information and further improve the

387

hedging performance. The VHAR model of RCov in Panel II and the HAR model of

388

RMVHR in Panel I utilize exactly the same information set (intraday five-minute

389

returns of spot and futures) and have similar structures. Since the HAR model in Panel

390

I leads to better hedging performance in terms of all these four hedging performance

391

indicators, we conclude that constructing the RMVHR and directly forecasting it is a

392

more efficient way to utilize intraday information. Last but not the least, the HAR17

393

GARCH model in Panel I utilizes the same length of past RMVHR series as that of the

394

HAR model, using GARCH-skewed-t innovations instead of normal innovations. Since

395

these two models have similar HE, while the HAR-GARCH model has lower HR,

396

higher PR and higher SR than the HAR model, we conclude that modeling the variance

397

of the dynamic hedge ratio is valuable for improving the dynamic hedging performance.

398

6. Robustness Checks

399

To test the robustness of the above results to different market conditions, we use

400

the nonparametric change point model (NPCPM) (Ross et al. 2011) to detect the

401

different volatility regimes of the CSI 300 index in the forecast period. The NPCPM

402

detects the shifts in the volatility by sequential application of Mood’s test (Mood, 1954),

403

which is a nonparametric test for comparing the variances of two samples. Since the

404

Mood’s test assumes the independence of observations, we filter the original return

405

series using a GARCH(1,1) model with student-t innovations following Ross (2013),

406

and use the standardized residuals for the sequential Mood’s tests.

407 408

Assume the two samples for variance comparison are

(r

2,1

1,1

, r1,2 ,..., r1,a )

and

, r2,2 ,..., r2,b ) , where a+b=T. The Mood’s test statistic can be calculated as:

T + 1  ∑ i=1 rank ( r1,i ) − 2  , where rank r1,i 2

409 = M

(r

a

( )

is the rank of r1,i in the combined

410

sample of length T. By comparing the standardized Mood’s test statistic with the

411

simulated thresholds reported in Ross et al. (2011), we can decide whether the null

412

hypothesis of equal variance is rejected. The NPCPM applies sequential Mood’s tests

413

in the following manner to detect the volatility change points:

414

1) Divide the out-of-sample period into two contiguous samples. The first sample 18

415

contains the initial 22 (a month) observations, and the second sample contains

416

the remaining 953 (975-22=953) observations.

417

2) Perform the Mood’s test on these two samples.

418

3) If the null hypothesis of equal variance is not rejected, prolong the first sample

419

by 1 observation, and thus the second sample contains the remaining 952

420

observations. Perform the Mood’s test on these two updated samples.

421

4) Repeat procedure 3) until the null hypothesis is rejected, which means a

422

volatility change point has been detected. Flag this change point and repeat

423

procedures 1)-3) starting from the first observation after the change point.

424

[Insert Figure 3 Here]

425

Figure 3 displays the volatility regimes detected by the NPCPM in the out-of-

426

sample period from January 2, 2014 to December 29, 2017. There are three detected

427

volatility regimes. The first regime is from January 2, 2014 to November 3, 2014,

428

altogether 203 trading days. We refer to it as the low volatility regime (L) since the CSI

429

300 index is very tranquil during this period. The second regime is from November 4,

430

2014 to August 31, 2016 (448 trading days). We refer to it as the high volatility regime

431

(H) since the CSI 300 index is extremely volatile during this period. The last regime is

432

from September 1, 2016 to December 29, 2017 (324 trading days). We again refer to it

433

as the low volatility regime (L) due to its similarity with the first regime.

434

[Insert Table 4 - 5 Here]

435

We perform hedging performance comparison on each of these three volatility

436

regimes, and report the results in Tables 4-5. Comparing these two tables, we can see

437

that the hedging effectiveness is always lower during the low volatility regime than 19

438

during the high volatility regime, with the only exception of the NAÏVE method. This

439

observation confirms the appropriateness of our partition of volatility regimes.

440

Our insights from Table 4 for the low volatility regimes are completely consistent

441

with those from Table 3, and our insights from Table 5 for the high volatility regime are

442

also consistent with those from Table 3, which we summarize as follows:

443

1) The RMVHR-based models (Panel I) have much lower HR, higher HE, PR and

444

SR than those of the daily return based models (Pane II excluding the VHAR model) in

445

both the low volatility regimes and the high volatility regime.

446

2) The lowest hedge ratio, the highest hedging effectiveness, portfolio return and

447

Sharp ratio always come from the RMVHR-based models, regardless of the volatility

448

regime considered.

449

3) Incorporating the information in the realized covariance matrix into the DCC

450

model can effectively improve the dynamic hedging performance, regardless of the

451

volatility regime considered.

452

4) Directly modeling the realized covariance matrix with the VHAR model can

453

better utilize the intraday information than the DCC-RV model and further improve the

454

hedging performance, regardless of the volatility regime considered.

455

5) Directly forecasting the RMVHR is a more efficient way to utilize intraday

456

information during the low volatility regimes, since in Table 4 the HAR model of

457

RMVHR in Panel I has lower HR, higher HE, PR and SR than the VHAR model of

458

RCov in Panel II. However, this conclusion does not hold in the high volatility regime.

459

Nevertheless, by replacing the normal innovations in the HAR model with the GARCH20

460

skewed-t innovations, the HAR-GARCH model in Panel I has lower HR, higher PR

461

and SR than the VHAR model, regardless of the volatility regime considered.

462

7.

Concluding Remarks

463

This paper examines the incremental value of stock futures hedges in China’s

464

market by directly forecasting the realized minimum-variance hedge ratio calculated

465

from high-frequency prices of CSI 300 spot and futures over the conventional hedging

466

models.

467

Using five-minute data of China’s CSI 300 index and index futures, we made out-

468

of-sample forecasts of the RMVHR with a number of popular time-series models, and

469

used the forecasts to perform dynamic hedging of the CSI 300 index futures. We also

470

include the static OLS and ECM models, the VHAR model, the dynamic DCC model

471

based on daily returns, and the DCC-RV model using five-minute prices for comparison.

472

In addition, we detected three volatility regimes in the forecast period using the

473

nonparametric change point model (Ross et al. 2011). Using the hedge ratio, the

474

hedging effectiveness, the portfolio return and the Sharp ratio as criteria, we conducted

475

hedging performance comparison in the out-of-sample forecast period as well as in each

476

detected volatility regime.

477

Our results show that the dynamic hedging performance of CSI 300 index futures

478

using RMVHR dominates that of the conventional methods in terms of all these four

479

hedging performance indicators in the out-of-sample forecast period. Furthermore, the

480

superiority of the RMVHR-based methods is consistent during different volatility

481

regimes of China’s financial markets, including China’s abnormal market fluctuations

482

in 2015. Our research also shed some new lights on the conventional hedging models. 21

483

For instance, incorporating information in the realized measures from high-frequency

484

data improves the dynamic hedging performance. In addition, the VHAR model that

485

directly models the realized covariance matrix better utilizes the intraday information

486

and outperforms the DCC-RV model.

487

Our research is instructive for futures hedgers, risk managers and academics.

488

Future work would involve exploring forecast combination techniques to improve the

489

forecasting capability of RMVHR and the dynamic hedging performance.

490

Acknowledgements

491

This work was supported by the National Natural Science Foundation of China

492

(Research Grant No. 71671084); and the Specialized Research Fund for the Doctoral

493

Program of Higher Education of China (Research Grant No. 20120091120003).

494 495

22

496

References

497 498 499

Andersen, T. G., & Bollerslev, T. (1998). Answering the skeptics: Yes, standard volatility models do provide accurate forecasts. International Economic Review, 34(9): 885-905.

500 501

Baillie, R.T., Myers, R.J., 1991. Bivariate GARCH estimation of the optimal commodity futures hedge. Journal of Applied Econometrics 6, 109–124.

502 503 504

Barndorff-Nielsen, O. E., & Shephard, N. (2004). Econometric analysis of realised covariation: high frequency based covariance, regression and correlation in financial economics. Econometrica, 72(3): 885–925.

505 506 507 508

Barndorff-Nielsen, O. E., Hansen, P. R., Lunde, A. & Shephard, N. (2011). Multivariate realised kernels: Consistent positive semi-definite estimators of the covariation of equity prices with noise and non-synchronous trading. Journal of Econometrics, 162(2): 149-169.

509 510 511

Bollerslev, T. (1990). Modelling the coherence in short-run nominal exchange rates: a multivariate generalized ARCH model. Review of Economics & Statistics, 72(72):498-505.

512 513 514

Busch, T., Christensen, B. J., & Nielsen, M. Ø. (2011). The role of implied volatility in forecasting future realized volatility and jumps in foreign exchange, stock and bond markets. Journal of Econometrics, 160(1): 48-57.

515 516

Corsi, F. (2009). A simple approximate long-memory model of realized volatility. Journal of Financial Econometrics, 7(2): 174-196.

517 518

Ederington, L. H. (1979). The hedging performance of the new futures markets. Journal of Finance, 34(1):157-170.

519 520

Engle, R. F. (2002). Dynamic conditional correlation: a simple class of multivariate GARCH models. Journal of Business & Economic Statistics, 20(3): 339-350.

521 522

Engle, R. F., & Kroner, F. K. (1995). Multivariate simultaneous generalized ARCH. Econometric Theory, 11(1): 122-150.

523 524

Ghosh, A. (1993). Hedging with stock index futures: Estimation and forecasting with error correction model. The Journal of Futures Markets, 13(7):743–752.

525 526

Ghysels, E., Santa-Clara, P., & Valkanov, R. (2005). There is a risk-return trade-off after all. Journal of Financial Economics, 76: 509-548.

527 528 529

Han, L. Y., & Ren, G. Y. (2012). Hedging strategy with futures based on prediction of realized second moments: An application to stock index futures. Systems Engineering – Theory & Practice, 32(12): 2629-2636 (in Chinese).

530 531 532

He, P. and Yang, Z. 2012. Comparison Study of Hedge Effectiveness of Hang Sheng Stock Index and CSI 300 Stock Index Futures Contracts [J]. Review of Investment Studies, 4, p.014. 23

533 534 535

Hou, Y. and Li, S., 2013. Hedging performance of Chinese stock index futures: An empirical analysis using wavelet analysis and flexible bivariate GARCH approaches. Pacific-Basin Finance Journal, 24, pp.109-131.

536 537

Hsu, C.C., Tseng, C.P. and Wang, Y.H., 2008. Dynamic hedging with futures: A copulabased GARCH model. Journal of Futures Markets, 28(11), pp.1095-1116.

538 539

Johnson, L. L. (1960). The theory of hedging and speculation in commodity futures. The Review of Economic Studies, 27(3): 139-151.

540 541 542

Kroner, K. F., & Sultan, J. (1993). Time-varying distributions and dynamic hedging with foreign currency futures. Journal of Financial & Quantitative Analysis, 28(4):535-551.

543 544

Lai, Y. S., & Sheu, H. J. (2010). The incremental value of a futures hedge using realized volatility. The Journal of Futures Markets, 30(9):874-896.

545 546 547

Lai, Y., Chen, C.W. and Gerlach, R., 2009. Optimal dynamic hedging via copulathreshold-GARCH models. Mathematics and Computers in Simulation, 79(8), pp.2609-2624.

548 549 550

Markopoulou, C. E., Skintzi, V. D., & Refenes, A. N. (2016). Realized hedge ratio: Predictability and hedging performance. International Review of Financial Analysis, 45: 121-133.

551 552 553

Miao, H., Ramchander, S., Wang, T. and Yang, D., 2017. Role of index futures on China's stock markets: Evidence from price discovery and volatility spillover. Pacific-Basin Finance Journal, 44, pp.13-26.

554 555

Mood, A. M. (1954). On the asymptotic efficiency of certain nonparametric two-sample tests. The Annals of Mathematical Statistics, 25(3): 514-522.

556 557 558

Pan, Z. and Sun, X., 2014. Hedging Strategy Using Copula and Nonparametric Methods: Evidence from China Securities Index Futures. International Journal of Economics and Financial Issues, 4(1), p.107.

559 560 561

Park, S.Y., Jei, S.Y., 2010. Estimation and hedging effectiveness of time-varying hedge ratio: flexible bivariate GARCH approaches. Journal of Futures Markets 30 (1), 71– 99.

562 563

Ross, G. J. (2013). Modelling financial volatility in the presence of abrupt changes. Physica A: Statistical Mechanics and its Applications, 392(2): 350-360.

564 565

Ross, G. J., Tasoulis, D. K., & Adams, N. M. (2011). Nonparametric monitoring of data streams for changes in location and scale. Technometrics, 53: 379-389.

566 567

Stein, J. L. (1961). The simultaneous determination of spot and futures prices. American Economic Review, 51(5):1012-1025.

568 569 570

Wei, Y., Wang, Y. and Huang, D., 2011. A copula–multifractal volatility hedging model for CSI 300 index futures. Physica A: Statistical Mechanics and its Applications, 390(23), pp.4260-4272. 24

571 572 573

Wen, X., Wei, Y. and Huang, D., 2011. Speculative market efficiency and hedging effectiveness of emerging Chinese index futures market. Journal of Transnational Management, 16(4), pp.252-269.

574 575 576

Yang, C.Y., Jhang, L.J. and Chang, C.C., 2016. Do investor sentiment, weather and catastrophe effects improve hedging performance? Evidence from the Taiwan options market. Pacific-Basin Finance Journal, 37, pp.35-51.

577 578

Yang, W., Allen, D.E., 2004. Multivariate GARCH hedge ratios and hedging effectiveness in Australian futures markets. Accounting and Finance 45, 301–321.

579 580

Zhang, L. (2011). Estimating Covariation: Epps Effect, Microstructure Noise. Journal of Econometrics, 160(1): 33-47.

581 582 583

Zhang, X. and Shen Z., 2011. The Impact of Introduction of Stock Index Futures on China’s Stock Market Volatility: Empirical Analysis based on High-Frequency Data of HS300 index futures [J]. Review of Investment Studies, 10, p.013.

584

25

585 586 587

Figure 1. Time series plots of the log daily prices for the CSI 300 index and the CSI 300 index futures from January 4, 2012 to December 29, 2017.

588

26

589 590 591 592 593

Figure 2. Time series plots of the realized volatilities for the CSI 300 index and the CSI 300 index futures, and the realized covariance between the spot and the futures from January 4, 2012 to December 29, 2017.

27

594 595 596

Figure 3. Volatility regimes detected by the NPCPM in the out-of-sample period from January 2, 2014 to December 29, 2017.

597

598

28

599 600 601 602

603 604 605 606 607 608

Table 1. Descriptive statistics for the realized volatilities (RVtS and RVtF), the realized covariance (RCovtS,F), and the realized minimum-variance hedge ratio (RMVHRt) of the CSI 300 index and index futures from January 4, 2012 to December 29, 2017 RMVHRt RVtS RVtF RCovtS,F Mean 0.6600 1.6101 2.1515 1.3713 Standard Deviation 0.1724 3.3081 5.6002 3.4555 Skewness 0.1303 7.2779 10.7185 10.0274 Kurtosis 3.4574 76.4803 172.1786 146.6007 *** *** *** ADF -3.1386 -10.7050 -12.7407 -12.4644*** PP -30.4818*** -16.8493*** -17.3488*** -16.5660*** LB(5) 1007.8*** 2719.2*** 2627.6*** 2695.7*** LB(10) 1698.6*** 3847.8*** 3708.5*** 3668.4*** LB(20) 2891.2*** 5575.6*** 5015.1*** 5007.5*** Note: JB represents the Jarque-Bera normality test statistics, ADF represents the Augmented-Dickey-Fuller test statistics, PP represents the Phillips-Perron test statistics, LB(k) represents the Ljung-Box Q-statistics for kth order serial correlation, *** represents the significance level of 1%. The orders of magnitude for the mean and the standard deviation of RVtS, RVtF and RCovtS,F are 10-4.

29

609 610

Table 2. Diagnostic statistics for the regression residuals of the HAR model of RMVHR in the whole sample period from January 4, 2012 to December 29, 2017. ARCH(5) ARCH(10) ARCH(20) Skew Kurt JB *** *** ** 3.9680 0.2185 3.4579 23.9404*** 2.3323 1.6147

611

Note: ARCH(·) represents the ARCH-LM test statistics. JB represents the Jarque-Bera

612

normality test statistics. ** indicates significance at the 5% level, significance at the 1% level.

613 614 615 616 617 618 619 620 621 622 623 624

30

***

indicates

625 626 627 628

Table 3. Hedging performance comparison in the out-of-sample forecast period from January 2, 2014 to December 29, 2017.

Panel I: modeling the RMVHR RS ARMA ARMA-GARCH ARFIMA HAR HAR-GARCH HR 62.7216% 65.8045% 65.8096% 70.3667% 70.5335% 69.3108% HE 58.3634% 58.7287% 58.6469% 58.5888% 58.5556% 58.7288% PR 5.0476% 3.8319% 3.8302% 3.0697% 2.7218% 3.2610% SR 2.7654% 2.1591% 2.1583% 1.7720% 1.5711% 1.8568% Panel II: modeling the daily returns (covariance matrix) OLS ECM DCC DCC-RV VHAR NAÏVE 91.8041% 91.1134% 71.0610% 100.0000% HR 81.2190% 80.8292% 45.0022% 49.5666% 58.5016% 43.4252% HE 55.2131% 55.2960% 1.8086% -2.8180% -0.3087% 2.7077% 0.1082% PR 1.7645% 1.0355% -1.4666% -0.1835% 1.5693% 0.0521% SR 1.0111% 629 Note: HR represents the hedge ratio, HE represents the hedging effectiveness, PR 630 represents the portfolio return, SR represents the Sharp ratio. The lowest HR, the 631 highest HE, the highest PR and the highest SR are in bold. 632 633 634

31

635 636 637 638

639 640 641 642 643 644

Table 4. Hedging performance comparison in the low volatility regime from January 2, 2014 to November 3, 2014 (203 trading days), and from September 1, 2016 to December 29, 2017(324 trading days), total is 527 trading days. Panel I: modeling the RMVHR RS ARMA ARMAARFIMA HAR HAR-GARCH GARCH HR 62.7947% 65.8457% 65.8505% 69.8919% 69.7844% 68.3916% HE 58.0378% 58.3540% 58.3542% 58.3191% 58.3797% 58.3610% PR 4.5457% 4.5333% 4.5324% 3.6916% 3.8158% 4.2150% SR 5.5696% 5.8338% 5.8331% 4.9322% 5.1236% 5.5303% Panel II: modeling the daily returns (covariance matrix) OLS ECM DCC DCC-RV VHAR NAÏVE HR 83.1962% 82.9233% 107.3345% 99.5945% 71.2981% 100.0000% HE 54.1327% 54.1596% 37.4661% 44.9733% 58.1988% 44.7205% PR 2.6491% 2.6513% -1.5033% -0.0839% 3.2563% -0.0217% SR 3.8532% 3.8483% -1.8078% -0.1130% 4.4796% -0.0283% Note: HR represents the hedge ratio, HE represents the hedging effectiveness, PR represents the portfolio return, SR represents the Sharp ratio. The lowest HR, the highest HE, the highest PR and the highest SR are in bold.

32

645 646 647

Table 5. Hedging performance comparison in the high volatility regime from November 4, 2014 to August 31, 2016 (448 trading days)

Panel I: modeling the RMVHR ARMARS ARMA ARFIMA HAR HAR-GARCH GARCH HR 62.6356% 65.7562% 65.7615% 70.9253% 71.4147% 70.3921% HE 58.7464% 59.1696% 59.1696% 58.9614% 58.8569% 58.8339% PR 5.6381% 3.0067% 3.0041% 2.3382% 1.4349% 2.1387% SR 2.2156% 1.2122% 1.2112% 0.9642% 0.5914% 0.8704% Panel II: modeling the daily returns (covariance matrix) OLS ECM DCC DCC-RV VHAR NAÏVE 73.5352% 81.1366% 70.7822% 100.0000% HR 78.8931% 78.3659% 53.8673% 54.9698% 58.8577% 41.9014% HE 56.4839% 56.6327% 0.8172% -4.3645% -0.5731% 2.0624% 0.2611% PR 0.7240% 0.3312% -1.6232% -0.2440% 0.8516% 0.0885% SR 0.2936% 648 Note: HR represents the hedge ratio, HE represents the hedging effectiveness, PR 649 represents the portfolio return, SR represents the Sharp ratio. The lowest HR, the 650 highest HE, the highest PR and the highest SR are in bold.

33

Suggest Documents