Risk Management

Lecture 6: Multivariate GARCH Models

Chen Tong

SOE & WISE, Xiamen University

December 18, 2023

Chen Tong (SOE&WISE) Risk Management December 18, 2023 1 / 31

Vector-GARCH (VEC)

▶ Bollerslev et al. (1988) introduced the vector GARCH

specification for modeling the conditional covariance

vec (Σt ) = vec(C) + Avec (εt−1 εt−1 ) + Bvec (Σt−1 )

′

where C is a k by k positive definite matrix and both A and B

2 2
are k by k parameter matrices.

Chen Tong (SOE&WISE) Risk Management December 18, 2023 2 / 31

Bivariate VEC-GARCH

To understand the richness of the specification, consider the evolu-

tion of the conditional covariance in a bivariate model:

The evolution of the conditional variance of the first asset,

In practice it is very difficult to use the vector GARCH model since

finding general conditions on A and B which will ensure that Σt is
positive definite is difficult.

Chen Tong (SOE&WISE) Risk Management December 18, 2023 3 / 31

 ▶ The scalar version of the VEC has covariance dynamics that
evolve according to
′
Σt+1 = C + αεt εt + βΣt

where matrix C is positive-definite and α and β are positive

scalars.

Chen Tong (SOE&WISE) Risk Management December 18, 2023 4 / 31

 ▶ The model has k(k + 1)/2 + 2 parameters and so estimation is
not feasible when k is large. Engle and Mezrich (1996) show
that the scalar VEC can be "targeted," which allows that
intercept to be concentrated out using a simple moment-based
estimator. Assuming the α + β < 1,

E [Σt+1 ] = C + αE [εt εt ] + βE [Σt ]

′

Σ = C + αΣ + βΣ
C = Σ(1 − α − β)

where Σ is the unconditional covariance of εt .Σ can be

estimated using the usual moment estimator
T
−1 ′
Σ̂ = T ∑ εt εt ,
t=1

Chen Tong (SOE&WISE) Risk Management December 18, 2023 5 / 31

 ▶ And the parameters which determine the dynamics can be
estimated by maximizing the likelihood over only α and β
using the targeted likelihood,
ˆ ′
Σt+1 = Σ(1 − α − β) + αεt εt + βΣt

Chen Tong (SOE&WISE) Risk Management December 18, 2023 6 / 31

BEKK-GARCH

▶ The primary insight of the BEKK (Baba, Engle, Kraft and

Kroner) is that quadratic forms are positive semi-definite and
the sum of a positive semi-definite matrix and a positive
definite matrix is positive definite.

▶ The covariance in a BEKK GARCH( 1,1) model evolves

according to
′ ′ ′ ′
Σt = CC + Aϵt−1 ϵt−1 A + BΣt−1 B

where C is a k × k lower triangular matrix and A and B are

k × k parameter matrices.

Chen Tong (SOE&WISE) Risk Management December 18, 2023 7 / 31

 Consider a bivariate BEKK,
′
σ11,t σ12,t c 0 c 0
[ ] =[ 11 ] [ 11 ]
σ12,t σ22,t c12 c22 c21 c22
2 ′
a11 a12 ϵ1,t−1 ϵ1,t−1 ϵ2,t−1 a a12
+[ ][ 2 ] [ 11 ]
a21 a22 ϵ1,t−1 ϵ2,t−1 ϵ2,t−1 a21 a22
′
b11 b12 σ σ12,t−1 b b12
+[ ] [ 11,t−1 ] [ 11 ]
b21 b22 σ12,t−1 σ22,t−1 b21 b22

The conditional variance of the first asset is given by

2 2 2
σ11,t = c11 + a11 ϵ1,t−1 + 2a11 a12 ϵ1,t−1 ϵ2,t−1
2 2 2 2
+ a12 ϵ2,t−1 + b11 σ11,t−1 + 2b11 b12 σ12,t−1 + b12 σ22,t−1

Chen Tong (SOE&WISE) Risk Management December 18, 2023 8 / 31

Diagonal BEKK-GARCH

▶ Estimation of full BEKK models rapidly becomes difficult as

the number of assets grows since the number of parameters in
2
the model is (5k + k) /2, and so is usually only appropriate
for k ≤ 5. The diagonal BEKK partially addresses some of the
number of parameters by restricting A and B to be diagonal
matrices.

▶ The covariance in a diagonal BEKK-GARCH(1,1) model

evolves according to
′ ′ ′ ′
Σt = CC + Ãϵt−1 ϵt−1 Ã + B̃Σt−1 B̃

where C is a k × k lower triangular matrix and à and B̃ are

diagonal parameter matrices.

Chen Tong (SOE&WISE) Risk Management December 18, 2023 9 / 31

Diagonal BEKK-GARCH

▶ The conditional covariances in a diagonal BEKK evolve

according to

σij,t = c̃ij + ai aj ϵi,t−1 ϵj,t−1 + bi bj σij,t−1

th ′
where c̃ij is the ij element of CC .

Chen Tong (SOE&WISE) Risk Management December 18, 2023 10 / 31

Scalar BEKK-GARCH
▶ The scalar BEKK further restricts the parameter matrices to
be common across all assets, and is a particularly simple (and
restrictive) model.

▶ The covariance in a scalar BEKK-GARCH(1,1) model evolves

according to
′ 2 ′ 2
Σt = CC + a ϵt−1 ϵt−1 + b Σt−1

▶ The scalar BEKK has one further advantage: it can easily be

covariance targeted.
2 2 ˆ 2 2
Σt = (1 − a − b ) Σ + a ϵt−1 ϵt−1 + b Σt−1

Chen Tong (SOE&WISE) Risk Management December 18, 2023 11 / 31

Constant Conditional Correlation (CCC) GARCH

▶ Bollerslev (1990) assumed that all of the dynamics in the

conditional covariance are attributable to changes in the
conditional variances.
⎡
⎢ σ11,t ρ12 σ1,t σ2,t ρ13 σ1,t σ3,t . . . ρ1k σ1,t σk,t ⎤
⎥
⎢
⎢ ⎥
⎥
⎢
⎢ ρ σ σ σ22,t ρ23 σ2,t σ3,t . . . ρ2k σ2,t σk,t ⎥
⎥
⎢ 12 1,t 2,t ⎥
Σt = ⎢
⎢
⎢ 13 σ1,t σ3,t
⎢ ρ ρ23 σ2,t σ3,t σ33,t . . . ρ3k σ3,t σk,t ⎥
⎥
⎥
⎥
⎢
⎢ ⎥
⎥
⎢
⎢ ⋮ ⋮ ⋮ ⋮ ⋮ ⎥
⎥
⎢
⎢ ⎥
⎥
⎣ ρ1k σ1,t σk,t ρ2k σ2,t σk,t ρ3k σ3,t σk,t ... σkk,t ⎦

Chen Tong (SOE&WISE) Risk Management December 18, 2023 12 / 31

 ▶ Constant Conditional Correlation GARCH (Bollerslev, 1990)
decomposes the conditional covariance into k conditional
variances and the conditional correlation, which is assumed to
be constant,
Σt = Dt RDt
Dt is a diagonal matrix with the conditional standard deviation
th th
of the i asset in its i diagonal position.

⎡
⎢ σ1,t 0 0 ... 0 ⎤
⎥
⎢
⎢ ⎥
⎥
⎢
⎢ 0 σ2,t 0 ... 0 ⎥
⎥
⎢ ⎥
Dt = ⎢
⎢
⎢
⎢ 0 0 σ3,t ... 0 ⎥
⎥
⎥
⎥
⎢
⎢ ⎥
⎥
⎢
⎢ ⋮ ⋮ ⋮ ⋮ ⋮ ⎥
⎥
⎢
⎢ ⎥
⎥
⎣ 0 0 0 . . . σk,t ⎦

Chen Tong (SOE&WISE) Risk Management December 18, 2023 13 / 31

 ▶ The conditional variances are typically modeled using standard
GARCH (1, 1) models.
2
σii,t = ωi + αi ri,t−1 + βi σii,t−1

▶ The conditional correlation is constant

⎡
⎢ 1 ρ12 ρ13 ⋯ ρ1k ⎤
⎥
⎢
⎢ ⎥
⎢
⎢ ρ12 1 ρ23 ⋯ ρ2k ⎥
⎥
⎥
⎢ ⎥
R=⎢
⎢
⎢
⎢ ρ13 ρ23 1 ⋯ ρ3k ⎥
⎥
⎥
⎥
⎢
⎢ ⋮ ⎥⎥
⎢
⎢ ⋮ ⋮ ⋮ ⋮ ⎥
⎥
⎢
⎢ ⎥
⎣ ρ1k ρ2k ρ3k ⋯ 1 ⎥ ⎦

Chen Tong (SOE&WISE) Risk Management December 18, 2023 14 / 31

CCC-GARCH

▶ Bollerslev (1990) shows that the CCC GARCH model can be

estimated in two steps. The first fits k conditional variance
models (e.g. GARCH) and produces√ the vector of standardized
residuals ut where ui,t = ϵi,t / σ̂ii,t . The second step estimates
the constant conditional correlation using the standard
correlation estimator on the standardized residuals.

Chen Tong (SOE&WISE) Risk Management December 18, 2023 15 / 31

Dynamic Conditional Correlation (DCC)

▶ Dynamic Conditional Correlation extends CCC GARCH by

introducing simple, scalar BEKK-like dynamics to the
conditional correlations, and so R in the CCC is replaced with
Rt in the DCC (Engle, 2002).

Chen Tong (SOE&WISE) Risk Management December 18, 2023 16 / 31

DCC-GARCH

▶ The covariance in a dynamic conditional correlation GARCH

model evolves according to

Σt = Dt Rt Dt

where
∗ ∗
Rt = Qt Qt Qt

Qt = (1 − a − b)R + aut−1 ut−1 + bQt−1

′

− 12
Qt = (Qt ⊙ Ik )
∗

where ut is √ the k × 1 vector of standardized returns

(ui,t = ϵi,t / σ̂ii,t )

Chen Tong (SOE&WISE) Risk Management December 18, 2023 17 / 31

Factor GARCH
▶ Assume we have following factor structure for the return of
asset i:
K
Ri,t = µi + ∑ βi,k fk,t + εi,t
k=1

▶ We could define the return vector Yt :

Yt ≡ [R1,t , R2,t , ..., RN,t ]

′

µ ≡ [µ1 , µ2 , ..., µN ]
′

Ft ≡ [f1,t , f2,t , ..., fK ,t ]

′

then we have
Yt = µt + BFt

Chen Tong (SOE&WISE) Risk Management December 18, 2023 18 / 31

 ▶ Assume εi,t is independent of factors, then we have

R F ′
Σt = BΣt B + Ωt

where Ωt is the conditional covariance matrix of εt .

Chen Tong (SOE&WISE) Risk Management December 18, 2023 19 / 31

Case 1: factors are observable

▶ If the factors are observable (e.g. Fama-french factors), we

F
could first estimate Σt using some multi-variate GARCH
models introduced before.

▶ Second, we need to estimate the loading matrix B using OLS,

and then calculate the residuals εt .

Chen Tong (SOE&WISE) Risk Management December 18, 2023 20 / 31

Case 2: factors are unobservable
▶ If the factors are unobservable, we could estimate the factors
using principle components analysis (PCA).

▶ Assume that the factors are conditionally uncorrelated. (PCA

uses the unconditional covariance of returns and so only
guarantees that the factors are unconditionally uncorrelated.)

⎡
⎢
2
σf 1,t 0 0 ... 0 ⎤
⎥
⎢
⎢ ⎥
⎥
⎢
⎢ 2 ⎥
⎥
⎢
⎢
0 σf 2,t 0 ... 0 ⎥
⎥
Σt = ⎢
⎢ ⎥
⎥
f 2
⎢
⎢ 0 0 σf 3,t ... 0 ⎥
⎥
⎢
⎢ ⎥
⎥
⎢
⎢ ⋮ ⋮ ⋮ ⋮ ⋮ ⎥
⎥
⎢
⎢ ⎥
⎥
⎢
⎣ 0 0 0
2
. . . σfK ,t ⎥
⎦

▶ Modeling σfk,t using univarite GARCH model.

Chen Tong (SOE&WISE) Risk Management December 18, 2023 21 / 31

Conditional covariance of the residuals
▶ The conditional covariance of the residuals can be assumed to
be constant and diagonal,
⎡
⎢
2
σ1 0 0 ... 0 ⎤
⎥
⎢
⎢ ⎥
⎥
⎢
⎢ 2 ⎥
⎥
⎢
⎢
0 σ2 0 ... 0 ⎥
⎥
Ω=⎢
⎢
⎢
⎢ 0 0 σ3
2
... 0
⎥
⎥
⎥
⎥
⎢
⎢ ⎥
⎥
⎢
⎢ ⋮ ⋮ ⋮ ⋮ ⋮ ⎥
⎥
⎢
⎢ ⎥
⎥
⎢
⎣ 0 0 0
2
. . . σN ⎥
⎦

▶ Or diagonal and time-varying (univarite GARCH model):

⎡
⎢
2
σ1,t 0 0 ... 0 ⎤
⎥
⎢
⎢ ⎥
⎥
⎢
⎢ 2 ⎥
⎥
⎢
⎢
0 σ2,t 0 ... 0 ⎥
⎥
Ωt = ⎢
⎢
⎢
⎢ 0 0
2
σ3,t ... 0
⎥
⎥
⎥
⎥
⎢
⎢ ⎥
⎥
⎢
⎢ ⋮ ⋮ ⋮ ⋮ ⋮ ⎥
⎥
⎢
⎢ ⎥
⎥
⎢
⎣ 0 0 0
2
. . . σN,t ⎥
⎦

Chen Tong (SOE&WISE) Risk Management December 18, 2023 22 / 31

Advances in multivariate volatility modeling

▶ Note that we have following decomposition:

Σt = Dt Rt Dt

where
⎡
⎢ σ1,t 0 0 ... 0 ⎤
⎥
⎢
⎢ ⎥
⎥
⎢
⎢ 0 σ2,t 0 ... 0 ⎥
⎥
⎢ ⎥
Dt = ⎢
⎢
⎢
⎢ 0 0 σ3,t ... 0 ⎥
⎥
⎥
⎥
⎢
⎢ ⎥
⎥
⎢
⎢ ⋮ ⋮ ⋮ ⋮ ⋮ ⎥
⎥
⎢
⎢ ⎥
⎥
⎣ 0 0 0 . . . σk,t ⎦
and Rt is the correlation matrix.

▶ The key is to ensure the correlation matrix Rt is positive

definite.

Chen Tong (SOE&WISE) Risk Management December 18, 2023 23 / 31

Advances in multivariate volatility modeling
▶ In the bivariate case, we have

1 ρt
Rt = [ ]
ρt 1

so we could use the Fisher transformation

1 1+ρ
ρ ↦ F (ρ) ≡ log
2 1−ρ

which is a one-to-one mapping from (−1, 1) into R.

▶ Then we could model ρt in an unrestricted way: e.g.

2
F (ρt+1 ) = ω + βF (ρt ) + ut ut ∼ N(0, σ )

Chen Tong (SOE&WISE) Risk Management December 18, 2023 24 / 31

Extend the Fisher transformation

▶ How to extend the Fisher transformation into multivariate

case?

▶ Archakov and Hansen (2021, Econometrica) proposed a new

parametrization of correlation matrices, where positive
definiteness is an innate property.

▶ This parametrization can be viewed as a generalization of

Fisher transformation to higher dimensions.

Chen Tong (SOE&WISE) Risk Management December 18, 2023 25 / 31

Archakov and Hansen (2021, Econometrica)

▶ There is a unique spectral decomposition for correlation matrix

R:
′
⎛ λ1 0 ⋯ 0 ⎞⎛ ξ1 ⎞
⎜ ⎟ ⎜ ⎟
⎟⎜ ⎟
′
⋯ ξp ) ⎜ ⎟
0 λ2 0 ⋮ ⎜ ξ2 ⎟
R = ( ξ1 ξ2 ⎜
⎜ ⎟ ⎜ ⎟
⎜
⎜ ⋮ 0 ⋱ 0 ⎟⎜
⎟ ⎜
⎜ ⋮ ⎟
⎟
⎟
⎝ 0 ⋯ 0 λp ⎠⎝ ξp
′ ⎠

that is
N
′
R = ∑ λ k ξk ξk
k=1

and all λi are positive!

Chen Tong (SOE&WISE) Risk Management December 18, 2023 26 / 31

Archakov and Hansen (2021, Econometrica)

▶ Then the logarithm of correlation matrix is defined by:

N
log R = ∑ ξk ξk log(λk )
′

k=1

▶ And the new parametrization of correlation matrix is given by

γ(R) = vecl (log R)

where the vecl (⋅) extracts and vectorizes the elements below
the diagonal.

▶ Archakov and Hansen(2021) showed that γ(R) and R is a

one-to-one mapping.

Chen Tong (SOE&WISE) Risk Management December 18, 2023 27 / 31

 ▶ To illustrate this parametrization, consider following example

⎛⎡
⎢ 1.0 •
⎢
• ⎤
⎥
⎥ ⎞ ⎛ ⎡ ⎢
⎢
1.0 • • ⎤
⎥
⎥⎞
γ⎜⎢
⎢ ⎥
⎥ ⎟ ⎜ ⎢
⎢ ⎥
⎥⎟
⎜⎢
⎢ 0.8 1.0 • ⎥
⎥ ⎟ = vecl ⎜ ⎢
log ⎢ 0.8 1.0 • ⎥
⎥⎟
⎝⎢
⎢
⎣ 0.0 0.2 1.0
⎥
⎥
⎦⎠ ⎝ ⎢ ⎢
⎣ 0.0 0.2 1.0
⎥
⎥
⎦⎠

⎛⎡⎢
⎢
−0.53 • • ⎤ ⎥
⎥ ⎞ ⎡ ⎢
⎢
1.14 ⎤
⎥
⎥
⎜
= vecl ⎜⎢⎢
⎢ ⎥
⎥ ⎟ ⎢
⎢ ⎥
⎥
⎢ 1.14 −0.57 • ⎥
⎥ ⎟ = ⎢
⎢ −0.13 ⎥
⎥
⎝⎢⎢ ⎥
⎥ ⎠ ⎢
⎢ ⎥
⎥
⎣ −0.13 0.28 −0.03 ⎦ ⎣ 0.28 ⎦

Chen Tong (SOE&WISE) Risk Management December 18, 2023 28 / 31

 ▶ This parametrization is an extension of Fisher transformation:
2 1+ρ
1 ρ
1
log (1 − ρ ) 1
log 1−ρ
log ( )=( 2
1+ρ
2
2 )
ρ 1 1
2
log 1−ρ 1
2
log (1 − ρ )

and we have
1 ρ 1 1+ρ
γ ([ ]) = log
ρ 1 2 1−ρ

Chen Tong (SOE&WISE) Risk Management December 18, 2023 29 / 31

Properties of log R
▶ It is easy to verify that if R is a block correlation matrix, then
log R preserves the block structure.

⎛ 1.0 0.4 0.4 0.2 0.2 0.2 ⎞

⎜
⎜ 0.4 1.0 0.4 0.2 0.2 0.2 ⎟
⎟
⎜
⎜ ⎟
⎟
⎜ ⎟
R=⎜ ⎟
0.4 0.4 1.0 0.2 0.2 0.2
⎜
⎜ ⎟
⎟
⎜
⎜ 0.2 0.2 0.2 1.0 0.6 0.6 ⎟
⎟
⎜
⎜ ⎟
⎟
⎜
⎜ 0.2 0.2 0.2 0.6 1.0 0.6 ⎟
⎟
⎝ 0.2 0.2 0.2 0.6 0.6 1.0 ⎠

⎛ −0.16 0.349 0.349 0.104 0.104 0.104 ⎞

⎜
⎜ 0.349 −0.16 0.349 0.104 0.104 0.104 ⎟
⎟
⎜
⎜ ⎟
⎟
⎜ ⎟
log R = ⎜ ⎟
0.349 0.349 −0.16 0.104 0.104 0.104
⎜
⎜ ⎟
⎟ .
⎜
⎜ 0.104 0.104 0.104 −0.36 0.553 0.553 ⎟
⎟
⎜
⎜ ⎟
⎟
⎜
⎜ 0.104 0.104 0.104 0.553 −0.36 0.553 ⎟
⎟
⎝ 0.104 0.104 0.104 0.553 0.553 −0.36 ⎠

Chen Tong (SOE&WISE) Risk Management December 18, 2023 30 / 31

Applications: reducing dimensions
▶ For instance,

⎛ 0.68 ⎞
⎛ ⎡ ⎢
⎢
1.0 0.6 0.2 0.2 ⎤
⎥
⎥⎞ ⎜ ⎜
⎜
0.12 ⎟
⎟
⎟
⎜ ⎢
⎢ ⎥
⎥⎟ ⎜
⎜ ⎟
⎟
⎜
γ (R) = vecl ⎜ ⎢
⎢
0.6 1.0 0.2 0.2 ⎥
⎥⎟
⎟ ⎜
⎜
0.12 ⎟
⎟
⎜
⎜log ⎢ ⎥⎟ = ⎜ ⎟ = Aξ
⎜ ⎢ ⎢
⎢ 0.2 0.2 1.0 0.8 ⎥
⎥
⎥
⎟
⎟ ⎜
⎜
⎜ 0.12 ⎟
⎟
⎟
⎝ ⎢ ⎢
⎣ 0.2 0.2 0.8 1.0
⎥
⎥
⎦⎠ ⎜
⎜
⎜ 0.12
⎟
⎟
⎟
⎝ 1.08 ⎠

where
⎡
⎢ 1 0 0 0 0 0 ⎤
⎥ ⎛ 0.68 ⎞
⎢ ⎥
A =⎢
⎢ ⎥
⎥ ξ=⎜
⎜ 0.12 ⎟
′
⎢
⎢ 0 1 1 1 1 0 ⎥
⎥ , ⎟
⎢
⎢ ⎥
⎥ ⎝ 1.08 ⎠
⎣ 0 0 0 0 0 1 ⎦
So modeling γ (R) can be done through modeling ζ. The
parameters are reduced from 6 to 3 .

Chen Tong (SOE&WISE) Risk Management December 18, 2023 31 / 31