1 forecasting correlation and covariance with a range-based dynamic conditional correlation model...
TRANSCRIPT
1
Forecasting Correlation and Covariance with a Range-Based Dynamic
Conditional Correlation Model
Ray Y Chou Institute of Economics Academia Sinica
Nathan Liu National Chiao Tung University
Chun-Chou Wu Chung Yuan University
2
Outline
Introduction
Dynamic Conditional Correlation (DCC) Model
Methodology
Data Analysis
Empirical Analysis
Conclusion
3
Introduction
It is of primary importance in the practice of portfolio management, asset allocation and risk management to have an accurate estimate of the covariance matrices for asset prices.
The univariate ARCH/GARCH family of models provides effective tools to estimate the volatilities of individual asset prices, see Bollerslev, Chou and Kroner (1992), Engle (2004).
It is an active research issue in estimating the covariance or correlation matrices of multiple, including VECH model of Bollerslev, Engle and Wooldridge (1988), BEKK model of Engle and Kroner (1995) and the constant correlation model of Bollerslev (1990) among others.
4
Introduction (Cont.)
The constant correlation model is too restrictive as it precludes the dynamic structure of covariance to be completely determined by the individual volatilities.
The VECH and the BEKK models are more flexible in allowing time-varying correlations. Nevertheless the computational difficulty of the estimation of the model parameters has limited the application of these models to only small number of assets.
Engle and Sheppard (2001), Engle (2002a) and Engle, Cappiello and Sheppard (2003) provide a solution to this problem by using a model entitled the Dynamic Conditional Correlation Multivariate GARCH (henceforth DCC).
5
Introduction (Cont.)
Other estimation methods of time-varying correlation are proposed by Tsay (2002) and by Tse and Tsui (2002).
there is a growing awareness of the fact that the high/low range data of asset prices can provide sharper estimates and forecasts than the return data based on close-to-close prices. Studies of suppoting evidences include Parkinson (1980), Garman and Klass (1980), Wiggins (1991), Rogers and Satchell (1991), Kunitomo (1992), and more recently Gallant, Hsu and Tauchen (1999), Yang and Zhang (2000), Alizadeh, Brandt and Diebold (2001), Brandt and Jones (2002), Chou, Liu and Wu (2003) and Chou (2005a, 2005b).
6
Conditional Covariance/Correlation
)( ,2,11,12 tttt rrECOV
1 1, 2,12, 2 2
1 1, 1 2,
( )
( ) ( )
t t tt
t t t t
E r r
E r E r
Traditional Conditional Covariance and Correlation
Moving Average with a 100-week window, MA(100)
1
100,2,1,12 100
1 t
tssst rrCOV
))((
ˆ1
100
2,2
1
100
2,1
1
100,2,1
t
tss
t
tss
t
tsss
t
rr
rr
7
Conditional Covariance/Correlation
Exponentially-Weighted Moving Average (EWMA) model where the weights decrease exponentially
As to the coefficient is usually called the exponential smoother in this model by RiskMetrics.
1
1,2,1
1,12 )1(
t
sss
stt rrCOV
))((
ˆ1
1
2,2
11
1
2,1
1
1
1,2,1
1
t
ss
stt
ss
st
t
sss
st
t
rr
rr
8
Dynamic Conditional Correlation model
The multivariate GARCH model proposed assumes that returns from k assets are conditionally multivariate normal with zero expected value and covariance matrix Ht.
where Dt is the kxk diagonal matrix of time varying standard deviations from univariate volatility models with the square root of on the diagonal, and is the time varying correlation matrix.
),0(~1 ttt HNIr
tttt DRDH
ith tR tR
9
Dynamic Conditional Correlation model (Cont.)
It is important to recognize that although the dynamic of the Dt matrix has usually been structured as a standard univariate GARCH, it can take many other forms.
The standardized series is , hence
are the residuals standardized by their conditional standard deviation.
A general dynamic correlation structure proposed by Engle (2000a) is:
ttt rDZ 1
);0(~ tt RNZ
111 ')'( tttt QBZZABASQ 2/12/1 }{}{ tttt QdiagQQdiagR
10
Dynamic Conditional Correlation model (Cont.)
where S is the unconditional covariance of the standardized residuals. where i is a vector of ones and o is the Hadamard product of two identically sized matrices which is computed simply by element by element multiplication. It’s shown that if A , B, and (ii’-A-B) are positive semi-definite, then Q will be positive semi-definite.
The typical element of Rt will be of the form .
Engle and Sheppard (2001) show the necessary conditions for Rt to be positive definite and hence a correlation matrix with a real, symmetric positive semi-definite matrix, with ones on the diagonal.
jtit
ijt
qqq
11
Dynamic Conditional Correlation model (Cont.)
The loglikelihood of this estimator can be written:
Let the parameters in Dt be denoted and the additional parameters in Rt be denoted .
T
tttttt
T
ttttttttt
T
ttttt
ZRZRDk
rDRDrDRDk
rHrHkL
1
1
1
111
1
1
'loglog2)2log(2
1
'log)2log(2
1
'log)2log(2
1
12
Dynamic Conditional Correlation model (Cont.)
The log likelihood can written as the sum of a volatility part and a correlation part.
The volatility term is:
and the correlation component is:
),(, CV LLL
t
ttttV rDrDnL 22'log)2log(
2
1
tttttttC ZZZRZRL ''log
2
1, 1
13
Dynamic Conditional Correlation model (Cont.)
The volatility part of the likelihood is the sum of individual GARCH likelihood if Dt is determined by a GARCH specification
which can be jointly maximized by separately maximizing each term. If Dt is determined by a CARR specification,
where is the conditional standard deviation as computed from a scaled expected range from the CARR model.
t
k
i ti
titiV h
rhL
1 ,
2,
, )log()2log(2
1
t
k
i ti
titiV
rL
12*,
2,*
, )log(2)2log(2
1
*,ti
14
Dynamic Conditional Correlation model (Cont.)
The second part of the likelihood will be used to estimate the correlation parameters. As the squared residuals are not dependent on these parameters, they will not enter the first order conditions and can be ignored.
The two step approach to maximizing the likelihood is to find
and then take this value as given in the second stage,
VLmaxargˆ
,ˆmax CL
15
Methodology
For bivariate cases, we use the following GARCH and CARR model to perform the first step procedure in estimating the volatility parameters:
GARCH ( return-based conditional volatility model ) ,k=1,2
CARR ( range-based conditional volatility model ) ,k=1,2 ,where ,
tktkr ,, ),0(~| ,1, tkttk hNI
1,2,, tkkitkkktk hh
tktka
tk hrz ,,, /
tktk u ,, );1exp(~| 1, ttk Iu
1,1,, tkktkkktk *
,,, / tktkc
tk rz tkktk adj ,*
, k
kadj
ˆ
16
Methodology (Cont.)
In estimating the correlation parameters, we also allow two specifications of the dynamic structures in the correlation matrix.
MR_DCC (mean-reverting)
111 ')'( tttt QBZZABASQ
2
1,21,21,1
1,21,12
1,1
23
31
2233
3311
12
12
,2,12
,12,1
11
11
1
1
ttt
ttt
tt
tt
zzz
zzz
aa
aa
baba
baba
q
q
T
ttt zz
Tq
1,2,112
1
1,21,12
1,121,1
23
31
tt
tt
bb
bb
17
Methodology (Cont.)
I_DCC (integrated)
111 )1( tttt QAZZAQ
1,21,12
1,121,1
23
31
21,21,21,1
1,21,12
1,1
23
31
,2,12
,12,1
11
11
tt
tt
ttt
ttt
tt
tt
aa
aa
zzz
zzz
aa
aa
18
Methodology (Cont.)
The estimated conditional correlation coefficients are
Like volatilities, the correlation and covariance matrices are also unobservable. We use daily data to construct weekly realized covariance/correlation observations.
The purpose of this is to use these “measured covariance/ correlation” denoted MCOV/MCORR respectively as benchmarks in determining the relative performance of the return-based DCC model and the range-based DCC models.
tttt qqq ,2,1,12,12 /
19
Methodology (Cont.)
We perform the Mincer-Zarnowitz regressions for the in-sample comparison :
where
treturnttMCORR ,110 ˆ
trangettMCORR ,220 ˆ
tranget
returnttMCORR ,3210 ˆˆ
treturn
tt COVMCOV ,110
trange
tt COVMCOV ,220
trange
treturn
tt COVCOVMCOV ,3210
ttttCOV ,2,1 ˆˆˆ
20
Methodology (Cont.)
Following Engle and Sheppard (2001), the out-of-sample conditional correlations MR_DCC are computed.
For T+h, with , the correlation is
2,2,2,1
,2,12,1
23
31
2233
3311
12
12
1,21,12
1,121,1
11
11
1
1
TTT
TTT
TT
TT
zzz
zzz
aa
aa
baba
baba
q
q
TT
TT
bb
bb
,2,12
,12,1
23
31
1,21,11,121 / TTTT qqq
2233
3311
12
12
,2,12
,12,1
11
11
1
1
baba
baba
q
q
hThT
hThT
1,21,12
1,121,1
2233
3311
hThT
hThT
baba
baba
21
Data Analysis
The data employed in this study comparises 835 weekly observations on the S&P 500 Composite (henceforth S&P 500), NASDAQ stock market index and bond yield for 10 years time to maturities spanning the period 4 January 1988 to 2 January 2004. Two variant DCC structures to estimate the model’s significance are investigated.
Let MCORRt and MCOVt represent the realized multi-correlation coefficients and multi-covariances, respectively. The MCORRt is defined as:
where symbol the tradable days during the t week
1
1
i
ittMCORR
22
Data Analysis (Cont.)
is called the correlation coefficient at the i trading day of the week based on two indices that we have selected. We can use the daily returns data and fit them into the MR_DCC model, so that the can be computed. The multi-covariances (MCOVt) can be expressed as
Where represents the daily returns of k index at the i trading day during the t week.
it
it
121 )(
i
it
itt rrMCOV
iktr
23
Summary Statistics
Index S&P500 Nasdaq 10-year Tbond
Type range return range Type range return
Mean 3.136 0.182 4.204 Mean 3.136 0.182
Median 2.657 0.339 3.225 Median 2.657 0.339
Maximum 14.534 7.492 31.499 Maximum 14.534 7.492
Minimum 0.707 -12.330 0.514 Minimum 0.707 -12.330
Std. Dev. 1.730 2.153 3.224 Std. Dev. 1.730 2.153
Skewness 1.768 -0.458 2.411 Skewness 1.768 -0.458
Kurtosis 8.025 5.733 12.782 Kurtosis 8.025 5.733
Jarque-Bera 1312.227 288.695 4132.758 Jarque-Bera 1312.227 288.695
24
Return- Versus Range-based DCC Models In-sample MCORR Forecasting
Panel A: S&P500 and Nasdaq
MCORR 0̂ 1̂ 2̂ R-squared Return 0.251 (10.425) 0.707 (24.691) 0.409 Range 0.347 (16.937) 0.597 (24.539) 0.389 MR _DCC
BOTH 0.257 (10.524) 0.445 (7.435) 0.258 (5.277) 0.425 Return 0.250 (7.806) 0.705 (18.221) 0.316 Range 0.280 (10.676) 0.672 (21.554) 0.350 I_DCC
BOTH 0.171 (5.689) 0.363 (6.545) 0.440 (10.122) 0.392 MA100 0.259 (9.421) 0.699 (21.22) 0.330
Exp. Smoothing 0.375 (20.745) 0.555 (25.732) 0.489
Panel B: S&P500 and Tbond Return 0.030 (4.524) 0.848 (57.328) 0.777 Range 0.023 (3.789) 0.860 (70.081) 0.800 MR _DCC
BOTH 0.021 (3.621) -0.399 (-3.513) 1.256 (11.066) 0.803 Return 0.023 (3.539) 0.965 (59.105) 0.794 Range 0.007 (1.164) 0.945 (72.435) 0.811 I_DCC
BOTH 0.008 (1.458) 0.128 (1.161) 0.822 (7.730) 0.811 MA100 -0.017 (-1.704) 0.890 (36.625) 0.587
Exp. Smoothing 0.025 (3.947) 0.846 (67.370) 0.792
Panel C: Nasdaq and Tbond Return 0.026 (5.551) 0.851 (57.019) 0.777 Range 0.029 (6.574) 0.818 (67.500) 0.816 MR _DCC
BOTH 0.031 (7.113) -0.437 (-5.573) 1.222 (16.445) 0.821 Return 0.032 (6.888) 0.955 (57.549) 0.792 Range 0.018 (3.987) 0.890 (60.008) 0.802 I_DCC
BOTH 0.022 (4.970) 0.351 (3.944) 0.572 (6.777) 0.807 MA100 -0.017 (-2.707) 0.871 (42.423) 0.658
Exp. Smoothing 0.030 (6.283) 0.758 (62.599) 0.776
25
Correlation fitting (S&P500 and Bond)
-.2
-.1
.0
.1
.2
.3
.4
.5
.6
.7
25 50 75 100 125 150 175 200
CORR_SP_BONDRHO_MR_ARHO_MR_C
28
Return- Versus Range-based DCC Models In-sample MCOV Forecasting
Panel A: S&P500 and Nasdaq MCOV 0 1 2 R-squared
Return 0.753 (1.968) 0.939 (10.178) 0.223 Range 0.339 (0.750) 0.867 (9.438) 0.354 MR _DCC
BOTH 0.677 (1.829) -0.212 (-1.209) 0.998 (5.553) 0.357 Return 1.148 (2.962) 0.863 (9.571) 0.202 Range 0.384 (0.837) 0.848 (9.237) 0.351 I_DCC
BOTH 0.693 (1.836) -0.186 (-1.206) 0.961 (5.782) 0.354 MA100 0.788 (1.906) 1.011 (10.562) 0.210
Exp. Smooth 1.303 (3.577) 0.862 (9.855) 0.200 CCC 0.669 (1.690) 0.972 (9.994) 0.225
Panel B: S&P500 and Tbond Return -0.239 (-1.654) 0.960 (10.411) 0.230 Range -0.086 (-0.676) 0.860 (10.212) 0.309 MR _DCC
BOTH 0.007 (0.056) -0.559 (-2.025) 1.262 (4.864) 0.320 Return -0.298 (-2.014) 1.102 (10.850) 0.231 Range -0.199 (-1.489) 0.964 (10.312) 0.309 I_DCC
BOTH -0.148 (-1.139) -0.428 (-1.517) 1.262 (4.775) 0.315 MA100 -0.485 (-2.843) 1.070 (8.285) 0.177
Exp. Smooth -0.203 (-1.431) 0.945 (10.040) 0.227 CCC 2.527 (9.138) -3.982 (0.103) 0.103
Panel C: Nasdaq and Tbond Return -0.280 (-1.633) 0.934 (9.682) 0.163 Range -0.025 (-0.149) 0.702 (9.749) 0.220 MR _DCC
BOTH 0.007 (0.043) -0.301 (-1.157) 0.879 (4.418) 0.222 Return -0.264 (-1.556) 1.064 (9.921) 0.168 Range -0.171 (-1.004) 0.782 (9.847) 0.218 I_DCC
BOTH -0.174 (-1.030) -0.107 (-0.412) 0.844 (4.259) 0.218 MA100 -0.477 (-2.317) 0.955 (8.723) 0.152
Exp. Smooth -0.280 (-1.661) 0.814 (9.338) 0.161 CCC 1.696 (5.711) -14.546 (-5.789) 0.078
29
Return- versus Range-based DCC Models Out-of-Sample MCORR Forecasting
Panel A: S&P500 and Nasdaq
R-squared MR_DCC
Forecast horizon return-based range-based MA100
1 0.310 0.548 0.044 2 0.279 0.474 0.076
3 0.248 0.407 0.109
4 0.199 0.325 0.153
Panel B: S&P500 and Tbond
R-squared MR_DCC
Forecast horizon return-based range-based MA100
1 0.491 0.631 0.165
2 0.441 0.547 0.134
3 0.306 0.463 0.098
4 0.229 0.388 0.065
Panel C: Nasdaq and Tbond
R-squared MR_DCC
Forecast horizon return-based range-based MA100
1 0.593 0.786 0.443
2 0.673 0.756 0.397
3 0.705 0.706 0.348
4 0.719 0.654 0.299
30
Return- versus Range-based DCC Models Out-of-Sample MCOV Forecasting
Panel A: S&P500 and Nasdaq
R-squared MR_DCC
Forecast horizon return-based range-based MA(100) CCC
1 0.000 0.115 0.000 0.004
2 0.003 0.081 0.004 0.000
3 0.008 0.043 0.008 0.003
4 0.018 0.015 0.018 0.011
Panel B: S&P500 and Tbond
R-squared MR_DCC
Forecast horizon return-based range-based MA(100) CCC
1 0.001 0.136 0.001 0.006 2 0.000 0.070 0.000 0.010
3 0.000 0.079 0.002 0.004
4 0.007 0.046 0.006 0.006
Panel C: Nasdaq and Tbond
R-squared MR_DCC
Forecast horizon return-based range-based MA(100) CCC
1 0.009 0.114 0.053 0.003
2 0.002 0.083 0.032 0.007
3 0.004 0.066 0.032 0.000
4 0.004 0.046 0.025 0.000
31
Conclusion
A new estimator of the time-varying correlation/covariance matrices is proposed utilizing the range data by combining the CARR model of Chou (2005a) and the framework of Engle (2002)’s DCC model.
We find consistent results that the range-based DCC model outperforms the return-based models in estimating and forecasting covariance and correlation matrices, both in-sample and out-of-sample.
It can be applied to larger systems in a manner similar to the application of the return-based DCC model structures that is demonstrated in Engle and Sheppard (2001).
Future research will be useful in adopting more diagnostic statistics or tests based on value at risk calculations as is proposed by Engle and Manganelli (1999).