“Common Features in London”
Cass Business School, 16-17 December 2004
Elliptical Distributions and Dynamic
Conditional Correlation Models
Juan P. Cajigas and Giovanni Urga
Centre for Econometric Analysis (CEA@Cass) Cass Business School, London
Outline of the presentation
• DCC/ADCC models and the implication of the assumption of normality
• Elliptical distributions and the DCC model
• An empirical application using FTSE-All World • Some further theoretical developments:
– Misspecification of the pdf and Newey-Steigerwald (1997)
– Preliminary Monte Carlo evidence
Dynamic Conditional Correlation (S,L)
[
Engle (2002), Engle and Sheppard (2001)]
t t t t t t t D R D H H N I r = − ) , 0 ( ~ / 1 1 * 1 * ) ( ) ( − − = t t t t Q Q Q R
∑
∑
∑
∑
= − = − − = = + + ⎟ ⎠ ⎞ ⎜ ⎝ ⎛ − − = S s s t s L l l t l t l S s s L l l t Q Q Q 1 1 1 1 ' 1 α β α ε ε β ] ' [ t t E Q = ε ε position. diagonal i its on of element diagonal i the of root square with the matrix, diagonal a is th * th * t t Q QExtensions
[
Cappiello, Engle and Sheppard (2003)]
Asymmetric DCC (ADCC, also ADDCC)
where
A, B, and G are diagonal parameter matrixes
(
Q A Q A B Q B G NG)
A A B Q B G n n G Qt = − ' − ' − ' + 'εt−1ε't−1 + ' t−1 + ' t−1 't−1 t t t I n = [ε < 0]Dε ] [ntnt' E N =Further Models
Four special cases:
• CCC model (Bollerslev 1990): A=B=G=[0] • DCC (1,1) model (Engle 2002): • A(symmetric) DCC (1,1): • G(eneralized) Diagonal DCC (1,1): G=[0]
[ ]
, [ ][ ]
, ] [ ], 0 [ A a a B b b G = = ij = = ij =[ ]
, [ ][ ]
, [ ][ ]
, ] [g g A a a B b b G = ij = = ij = = ij =Extensions
[
Hafner and Franses (2003)]
(Generalized) Diagonal DCC (1,1) model:
the variance targeting approach is sacrificed by replacing with in order to assure that Qt will be positive-definite
(
a b)
A A B Q BQ
Qt = 1− 2 − 2 + 'εt−1ε 't−1 + ' t−1
Implications of the assumption of
normality
• Normality-MLE/QMLE = feasible + consistent but
inefficient DCC coefficients (Bollerslev and Wooldridge, 1992)
• Normality is not a satisfactory property for financial time series.
• Non normal distribution to achieve efficiency with implication for the first stage
Implications of the assumption of
normality
Efficiency loss in volatility parameters (univariate mode, % changes)
Implications of the assumption of
normality
Efficiency loss in correlation parameters
Alternative I:
Semiparametric (sp) estimation
• Univariate case:
Engle and González-Rivera (1991), Drost and Klaassen (1997), González-Rivera (1997),González-Rivera and Drost (1999).
• Multivariate case:
Hafner and Rombouts (2004)Remarks:
– More efficient than QMLE
but
– Estimation and inference quite difficult, and only feasible with small dimensions
Alternative II:
Assuming tick distributions
• Univariate case:
– Bollerslev (1987): Student-t distribution
– Baillie and Bollerslev (1989): Student-t and exponential-power distributions
– Nelson (1991): exponential-power distribution
– Verhoeven and McAleer (2003): Asymmetric Student-t,
asymmetric generalized error, asymmetric generalized Student-t, Gram-Charlier, and Pearson Type IV distribution.
• Multivariate case:
– Fiorentini, Sentana, Calzolari (2003, JBES): Multivariate Student-t in MGARCH
The main contribution of this paper:
elliptical-ADDCC (1,1)
We propose an ADDCC (1,1) model and its nested versions, using elliptical distributions for the
vector of standardized residuals.
We consider three distributions:
• Multivariate Laplace • Multivariate Student-t
Elliptical distributions
• Multivariate Laplace
• Multivariate Student-t
Two-Step estimation: feasible (1…)
Two-Step estimation: feasible (…2)
• Normal case
• Engle (2002): uses Newey-McFadden (1994, HoE) results on GMM to justify the use of MLE for consistency
Two-Step estimation: feasible
Two-step estimation
No feasible: lack identification of the degree of freedom parameter v
Two-Step estimation
No feasible: complexity of the functional form
Empirical application
Data
FTSE All-World weekly indices converted to US
denominated returns for 21 countries from 31/12/1993 to 09/04/2004 (T=538 weekly observations).
The countries are:
Australia Germany Netherlands Switzerland
Austria Hong Kong New Zealand United Kingdom Belgium Ireland Norway United States Canada Italy Singapure
Denmark Japan Spain France Mexico Swedem
Empirical results
• Six models were estimated: – Normal-DCC (1,1).
– Normal-DDCC (1,1)
– Normal-ADDCC (1,1) still running – Laplace-DCC(1,1)
– Laplace-DDCC (1,1)
Empirical results
• Step 1: Univariate volatilities were estimated with GARCH (1,1) models (Vol. parameters)
• Step 2: Estimation of the DCC and its variants (Corr. Parameters) using results from Step 1
• Table: note the trade-off between flexibility and feasibility
Model Vol. Parameters Corr. Parameters Total
DCC 63 2 65
DDCC 63 42 105
ADDCC 63 63 126
Empirical results
Variance parameters for the Normal Models
Australia s.e. Hong Kong s.e. Normway s.e.
omega 2.23E-05 1.02E-05 3.73E-06 2.52E-06 4.42E-06 2.85E-06
alpha 0.14605 0.067485 0.081495 0.034236 0.064193 0.025001
beta 0.69088 0.11078 0.90827 0.032557 0.91538 0.032557
Austria s.e. Ireland s.e. Singapure s.e.
omega 5.77E-06 3.00E-06 1.96E-05 9.60E-06 5.12E-06 4.14E-06
alpha 0.15053 0.042651 0.16321 0.078719 0.09062 0.038479
beta 0.82918 0.039649 0.73073 0.09833 0.89841 0.036273
Belgium s.e. Italy s.e. Spain s.e.
omega 1.89E-05 7.96E-06 0.00028691 0.000111 1.07E-05 5.77E-06
alpha 0.14596 0.042945 0.17492 0.072041 0.11659 0.03723
beta 0.74157 0.071355 0.20017 0.25074 0.83832 0.04905
Canada s.e. Japan s.e. Sweden s.e.
omega 1.62E-06 2.00E-06 5.94E-06 3.99E-06 6.78E-06 4.39E-06
alpha 0.11095 0.032954 0.090254 0.033284 0.11856 0.044855
beta 0.88905 0.033239 0.88638 0.040049 0.87031 0.038701
Denmark s.e. Mexico s.e. Switzerland s.e.
omega 5.31E-07 6.33E-07 9.05E-06 5.57E-06 3.78E-05 1.58E-05
alpha 0.059524 0.020538 0.079256 0.026151 0.18951 0.079165
beta 0.94047 0.019624 0.90382 0.029507 0.52119 0.15448
France s.e. Netherlands s.e. U.K. s.e.
omega 5.12E-06 3.22E-06 5.90E-06 2.88E-06 1.88E-06 1.43E-06
alpha 0.098613 0.034479 0.14257 0.043979 0.070166 0.024429
beta 0.87295 0.041849 0.82979 0.043862 0.91492 0.028555
Germany s.e. New Zeland s.e. USA s.e.
omega 1.90E-06 1.42E-06 1.47E-05 6.98E-06 7.41E-07 7.16E-07
alpha 0.11917 0.036376 0.12868 0.043759 0.093355 0.027173
Empirical results
Variance parameters for the Laplace Models
Australia s.e. Hong Kong s.e. Norway s.e. omega 2.25E-05 1.19E-05 5.70E-06 3.99E-06 7.83E-06 4.49E-06
alpha 0.13858 0.069143 0.075956 0.03585 0.077512 0.031291
beta 0.74244 0.10637 0.92361 0.03233 0.89433 0.040378
Austria s.e. Ireland s.e. Singapore s.e. omega 6.72E-06 3.45E-06 4.30E-06 2.83E-06 7.04E-06 4.99E-06
alpha 0.16198 0.047779 0.05587 0.029179 0.085002 0.039688
beta 0.833 0.040301 0.926 0.030964 0.90501 0.037034
Belgium s.e. Italy s.e. Spain s.e. omega 1.90E-05 8.31E-06 0.00015011 4.04E-05 1.73E-05 8.90E-06
alpha 0.17791 0.051261 0.16397 0.054078 0.14404 0.049907
beta 0.75896 0.065477 0.15052 0.15538 0.82288 0.057375
Canada s.e. Japan s.e. Sweden s.e. omega 3.94E-06 3.16E-06 4.70E-06 3.77E-06 7.38E-06 4.71E-06
alpha 0.11268 0.039156 0.082615 0.031753 0.10864 0.045311
beta 0.88732 0.040363 0.91738 0.031572 0.89136 0.036487
Denmark s.e. Mexico s.e. Switzerland s.e. omega 1.22E-06 1.01E-06 9.34E-06 6.01E-06 4.99E-05 2.36E-05
alpha 0.050246 0.020454 0.084133 0.02812 0.19755 0.09165
beta 0.94975 0.01951 0.91056 0.028187 0.49496 0.19161
France s.e. Netherlands s.e. U.K. s.e. omega 2.20E-06 1.90E-06 5.16E-06 2.72E-06 2.15E-06 1.69E-06
alpha 0.060811 0.023445 0.12643 0.042101 0.076305 0.028137
beta 0.93919 0.022732 0.86793 0.036157 0.92318 0.027638
Germany s.e. New Zeland s.e. USA s.e. omega 3.23E-06 1.97E-06 2.09E-05 9.41E-06 2.14E-06 1.17E-06
alpha 0.087564 0.033901 0.15096 0.053489 0.087155 0.030793
Empirical results
• Correlation parameters from the DCC (1,1) model
Parameters Normal DCC s.e. Laplace DCC s.e.
α 0.011518 0.002333 0.078904 0.01791 β 0.90327 0.027541 0.92109 0.023458
Empirical results
• Correlation parameters from the Normal-DDCC (1,1) model
Australia s.e. Hong Kong s.e. Norway s.e.
alpha 0.00896 0.00983 0.01723 0.01115 0.00675 0.00789
beta 0.90668 0.04939 0.85866 0.06246 0.91990 0.10491
Austria s.e. Ireland s.e. Singapore s.e.
alpha 0.01447 0.01263 0.02815 0.02184 0.00571 0.00755
beta 0.88716 0.13879 0.94368 0.05519 0.89168 0.06822
Belgium s.e. Italy s.e. Spain s.e.
alpha 0.05449 0.02645 0.06467 0.03358 0.03411 0.02412
beta 0.90249 0.03864 0.93533 0.04094 0.93180 0.03711
Canada s.e. Japan s.e. Sweden s.e.
alpha 0.02123 0.01600 0.00287 0.00543 0.03512 0.02200
beta 0.90909 0.05732 0.89469 0.07912 0.92594 0.03202
Denmark s.e. Mexico s.e. Switzerland s.e.
alpha 0.00725 0.00756 0.01026 0.01144 0.05763 0.03640
beta 0.92246 0.04624 0.93240 0.02613 0.91009 0.06829
France s.e. Netherlands s.e. U.K. s.e.
alpha 0.08221 0.02968 0.06931 0.02700 0.05608 0.03024
beta 0.91778 0.03317 0.92375 0.02804 0.90111 0.07283
Germany s.e. New Zeland s.e. USA s.e.
alpha 0.06568 0.02650 0.00219 0.00355 0.05173 0.03139
Empirical results
• Correlation parameters from the Laplace-DDCC (1,1) model
Australia s.e. Hong Kong s.e. Normway s.e.
alpha 0.0503 0.0303 0.3122 0.3727 0.4333 0.5545
beta 0.9497 0.2975 0.6878 0.0978 0.5667 0.2204
Austria s.e. Ireland s.e. Singapure s.e.
alpha 0.2312 0.5029 0.4022 0.3673 0.3202 0.5580
beta 0.7153 0.4389 0.5978 0.2403 0.6798 0.2279
Belgium s.e. Italy s.e. Spain s.e.
alpha 0.3278 0.4142 0.2212 0.0410 0.4962 0.3361
beta 0.6723 0.1392 0.7788 0.4702 0.5038 0.2047
Canada s.e. Japan s.e. Sweden s.e.
alpha 0.3171 0.5621 0.3671 0.3601 0.1653 0.1799
beta 0.6829 0.0481 0.6329 0.0834 0.8347 0.3087
Denmark s.e. Mexico s.e. Switzerland s.e.
alpha 0.2929 0.5182 0.2559 0.0635 0.2001 0.0379
beta 0.7072 0.4522 0.7441 0.3817 0.7999 0.3083
France s.e. Netherlands s.e. U.K. s.e.
alpha 0.3258 0.3145 0.4515 0.5771 0.3812 0.4881
beta 0.6742 0.0086 0.5485 0.2140 0.6188 0.2206
Germany s.e. New Zeland s.e. USA s.e.
alpha 0.2879 0.3050 0.2668 0.2972 0.2588 0.0325
Empirical results:
Main findings from the empirical
application
• Higher persistence in univariate variances
when the Laplace distribution is used
instead of the normal one
• The size of standard errors varies across
models.
• High levels of persistence in the correlation
under both distributions
Robust Conditional Moment Test.
Kroner and Ng (1998) in the framework of Wooldridge (1998):
ut is a “generalized residual” given by the distance
between a point of the scatter plot of and a corresponding point in the news impact surface.
If then uijt should be uncorrelated with any
variable known at time t-1 and the model will be correct.
Empirical results: diagnostic test
ijt jt it t h u = ε ε − 0 ) ( 1 = − ijt t u E jt itε ε
10 indicator variables are established, as in the sign and size bias tests of Engle and Ng (1993):
Wooldridge (1990):
where is the residual from the regression,
and is the vector of parameters in the MGARCH model. Under general regularity conditions the C test has an asymptotic distribution
Testing
[
]
∑
∑
= − = − = T t ijt mt T t ijt mt u T u C 1 2 1 2 2 1 1 λ λ∑
= − − ∂ + ∂ + = K k mt k ijt k mt h x 1 1 0 1 α α θ λ 10 ,..., 1 , 1 = − m mt λ ) ,..., (θ1 θK θ = ) 1 ( 2 χEmpirical results: tests
Indicator Normal-DCC Laplace_DCC Normal-DDCC Laplace_DDCC Normal-ADDCC Laplace-ADDCC
x1 22.53% 21.46% 9.72% 18.94% 13.70% 11.02% x2 17.05% 14.30% 7.57% 11.92% 8.63% 5.81% x3 18.86% 16.36% 8.08% 14.46% 9.33% 6.76% x4 25.06% 24.28% 10.87% 21.90% 14.12% 11.77% x5 1.37% 2.19% 1.06% 2.32% 1.73% 1.76% x6 1.79% 2.16% 1.66% 1.88% 2.18% 1.95% x7 3.92% 3.07% 1.31% 1.59% 2.38% 1.42% x8 2.91% 2.20% 1.48% 1.78% 1.49% 1.27% x9 3.14% 2.08% 1.61% 1.54% 2.41% 1.12% x10 3.79% 3.13% 1.84% 3.02% 2.07% 1.91%
Main findings from testing
• DCC model: Laplace outperforms
Normality
• DDCC model: Normality outperforms
Laplace
• ADDCC: Laplace outperforms Normality
Promising but too preliminary!!! Sample may be too short!!
First theoretical development:
misspecification of the pdf
• In the QMLE case we can obtain consistent but inefficient estimates (Bollerslev and Wooldridge, 1992), because of Gaussian hypothesis.
• QMLE might not be appropriate when we assume a non-Gaussian distribution (Newey and
Steigerwald, 1997, and Straumman, 2004) which is not the correct one.
Misspecification of the pdf
[
Intuition for the consistency of QMLE]• Only for the case of the Gaussian distribution the scale and location parameters coincide with the variance and mean parameters of the stochastic process
.
• For a density g(y) the natural location parameter µ and scale parameter σ are those that maximize,
⎥ ⎦ ⎤ ⎢ ⎣ ⎡ ⎟ ⎠ ⎞ ⎜ ⎝ ⎛ − + − σ µ σ g y E ln ln Y
When gY ( y)is Gaussian then µ = E(yt) and σ 2 = E(yt − µ)2 .
Newey and Steigerwald (1997)
Two main results:
1. If the true and assumed distributions of residuals are both symmetric then the identification
condition for consistency of the non-Gaussian QMLE is satisfied.
2. If this is not the case then an extra location parameter must be added to the standardized residuals
Monte Carlo experiments
• We design a Monte Carlo Experiment to analyze some of the theoretical results of Newey and
Steigerwald (1997)
• Because of the symmetric form of elliptical distributions we analyze only Result 1.
Conclusions
• The DCC model needs to be extended and
alternative distributions to be considered, more adequate than the normal to model financial time series.
• The fitting of the global equity markets improves when the Laplace distribution is used instead of the normal. Though the properties of the
estimators needs more analysis.
Future developments I
Consistency of the elliptical DCC model in cases of misspecification of the multivariate probability density function
The applicability and benefits of the
elliptical-DCC model should be evaluated via application to finance such VaR and asset allocation models.
Future developments II
• QMLE for conditional quantiles (Komunjer, 2004; Koenker and Bassett, 1978)
Asymmetric Laplace density and the corresponding QMLE reduces to non-linear quantile estimator
• Consistency as an identification issue (Wu, 1981)
Provides necessary conditions for asymptotic consistency for nonlinear models
• Use of robust procedures in estimating and testing (Hampel, 1974; Huber, 1981)