Dynamic Cholesky decomposition
CHAR models
Estimation
Asymptotics of Cholesky GARCH Models and Time-Varying Conditional Betas Serge Darolles, Christian Francq∗ and Sébastien Laurent CFE 2018 ∗ CREST
and university of Lille
Pisa · 14 December 2018 Supported by the ANR via the Project MultiRisk (ANR-16-CE26-0015-02)
Application
Dynamic Cholesky decomposition
CHAR models
Estimation
Application
Motivation Problem: Given some information set Ft−1 , it is often of interest to regress yt on the components of x t . Solution: yt − E (yt | Ft−1 ) = β 0y x,t {x t − E (x t | Ft−1 )} + ηt , with the dynamic conditional beta (DCB) β y x,t = Σ−1 xx,t Σxy ,t . Practical implementation: An ARCH-type model for the conditional variance
Σxx,t
Σxy ,t
of Σy x,t Σyy ,t x t − E (x t | Ft−1 ) is needed. t = yt − E (yt | Ft−1 ) A Cholesky GARCH model directly specifies the DCB.
Dynamic Cholesky decomposition
CHAR models
Estimation
Notation
Let t = (1t , . . . , mt )0 be a vector of m ≥ 2 log-returns satisfying 1/2
t = Σt
(ϑ0 )η t ,
where (η t ) is iid (0, In ), Σt = Σt (ϑ0 ) = Σ(t−1 , t−2 , . . . ; ϑ0 ) > 0, and ϑ0 is a d × 1 vector.
Application
Dynamic Cholesky decomposition
CHAR models
Estimation
Application
Engle (2002) DCC √ Σt = D t R t D t = ρijt σiit σjjt , 1/2
1/2
where D t = diag(σ11t , . . . , σmmt ) contains the volatilities of the individual returns, and R t = (ρijt ) the conditional correlations. The time series model needs to incorporate the complicated constraints of a correlation matrix. One often takes R t = (diag Q t )−1/2 Q t (diag Q t )−1/2 where Q t = (1 − θ1 − θ2 )S + θ1 u t−1 u 0t−1 + θ2 Q t−1 , √ with u t = (u1t . . . umt )0 , uit = it / σiit , θ1 + θ2 < 1.
Dynamic Cholesky decomposition
CHAR models
Estimation
Application
Engle (2016) DCB
Assuming
xt yt
| Ft−1
µ Σ xt , xx,t ∼N µy Σy x,t t
Σxy ,t Σyy ,t
we have −1 yt | x t ∼ N µyt + Σy x,t Σ−1 xx,t (x t − µx t ), Σyy ,t − Σy x,t Σxx,t Σxy ,t ⇒ β y x,t = Σ−1 xx,t Σxy ,t can be obtained by first estimating a DCC GARCH model.
Dynamic Cholesky decomposition
CHAR models
Estimation
Application
Drawbacks of DCC-based DCB
1) The stationarity and ergodicity conditions of the DCC are not well known. 2) The correlation constraints are complicated. 3) The asymptotic properties of the QMLE are unknown. 4) The effects of the DCC parameters on β t are hardly interpretable. We now introduce a class of Cholesky GARCH (CHAR) models that avoids all these drawbacks.
Dynamic Cholesky decomposition
CHAR models
Estimation
Application
Cholesky Decomposition of Σ = Var () Letting v1 := 1 , we have 2 = `21 v1 + v2 = β21 1 + v2 , where β21 = `21 is the beta in the regression of 2 on 1 , and v2 is orthogonal to 1 . Recursively, we have
i =
i−1 X j=1
`ij vj + vi =
i−1 X
βij j + vi ,
for i = 2, . . . , m,
j=1
where vi is uncorrelated to v1 , . . . , vi−1 , and thus uncorrelated to 1 , . . . , i−1 . In general the order of the series matters: Order of the series
Dynamic Cholesky decomposition
CHAR models
Estimation
Application
Cholesky Decomposition of Σ = Var () In matrix form, = Lv
and
v = B,
where L and B = L−1 are lower unitriangular and G := var(v) is diagonal. We obtain the Cholesky decomposition Σ = LGL0 (see Pourahmadi, 1999). Conditioning on Ft−1 , Σt = Lt G t L0t and Σ−1 = B 0t G −1 t B t . We thus need t - a diagonal ARCH-type model for the factors vector v t - a time series model for Lt (or B t ), without particular constraint.
Dynamic Cholesky decomposition
CHAR models
Estimation
Application
Example: Σ = LGL0 = DRD, m = 2 Letting g1 = σ12 = var(1 ),
2 = `21 1 +v2 ,
σ22 = var(2 )
g2 = varv2 ,
and ρ = cor(1 , 2 ), we have
1
L= `21
0 , 1
g1 G= 0
g1
Σ= `21 g1
0 g2
,
`21 g1 `221 g1 + g2
σ1 D= 0
σ12
= ρσ1 σ2
0 σ2
,
1 ρ , R= ρ 1
ρσ1 σ2 σ22
.
Positivity constraints: g1 > 0 and g2 > 0 or σ12 > 0, σ12 > 0 and ρ2 < 1.
Dynamic Cholesky decomposition
CHAR models
Estimation
Application
Example: Σ = LGL0 , Σ−1 = B 0 G −1 B, m = 3
1
0
L= `21
1
`31
`32
0 0 , 1
g1
Σ= `21 g1 `31 g1
1
B= −β21 −β31
0 1 −β32
0 0 , 1
`21 g1 `221 g1 + g2 `21 `31 g1 + `32 g2
g1 G= 0 0
`31 g1
0
g2
0 ,
0
g3
`21 l31 g1 + `32 g2 . 2 2 `31 g1 + `32 g2 + g3
Remark: In Σ = DRD the constraints on the elements of R are ρ212 + ρ213 + ρ223 − 2ρ12 ρ13 ρ23 ≤ 1. In Σ = LGL0 there is no constraint on the `ij ’s.
0
Dynamic Cholesky decomposition
CHAR models
Estimation
Application
A general model for the factors Assume 1/2
v t = Gt
ηt ,
(η t ) iid (0, In ),
where G t = diag(g t ) follows a GJR-like equation g t = ω0 +
q n p o X X 2− A0i,+ v 2+ + A v + B 0j g t−j , 0i,− t−i t−i i=1
j=1
with positive coefficients and v 2+ = t
+ 2 0 + 2 , v1t , · · · , vmt
v 2− = t
− 2 0 − 2 . v1t , · · · , vmt
Dynamic Cholesky decomposition
CHAR models
Estimation
Application
Markovian representation of the factors 0 0 2−0 0 Letting z t = v 2+ , v , g t:(t−p+1) , t:(t−q+1) t:(t−q+1) 0 0 0 0 Υ−0 , 00 0 , 00 ht = ω 00 Υ+ , 0 , ω , ω t 0 t 0 (p−1)m , with m(q−1) (q−1)m 2+ 2− Υ+ Υ− and obvious notations, we t = diag η t t = diag η t rewrite the model as z t = ht + H t z t−1 , where, in the case p = q = 1,
+ + Υ+ t A01,+ Υt A01,− Υt B 01
− − − . Ht = Υ A Υ A Υ B 01,+ 01,− 01 t t t A01,+ A01,− B 01
Dynamic Cholesky decomposition
CHAR models
Estimation
Stationarity of the factors
In view of z t = ht + H t z t−1 , there exists a stationary and ergodic sequence (v t ) satisfying 1/2
v t = Gt
η t if and only if γ0 = inf
t≥1
1 E(log kH t H t−1 . . . H 1 k) < 0. t
Application
Dynamic Cholesky decomposition
CHAR models
Estimation
Application
Stationarity of β t := −vech0 B t If (v t ) is stationary and ergodic (γ0 < 0), and ( det I m0 −
s X
) C 0i z
i
6= 0 for all |z| ≤ 1,
i=1
then βt = c0
1/2 1/2 v t−1 , . . . , v t−r , g t−1 , . . . , g t−r
+
s X
C 0j β t−j .
j=1
defines a stationary and ergodic sequence (and thus the existence of a stationary CHAR model).
Dynamic Cholesky decomposition
CHAR models
Estimation
Existence of moments If in addition Ekη 1 k2k1 < ∞
and
1 %(EH ⊗k 1 ) < 1,
for some integer k1 > 0, and kc 0 (x) − c 0 (y)k ≤ K kx − yka for some constants K > 0 and a ∈ (0, 1], then the CHAR model satisfies E k1 k2k1 < ∞.
Application
Dynamic Cholesky decomposition
CHAR models
Estimation
Application
A simpler triangular parameterization A tractable submodel is i 2 X 2 (k ) − α0i vk2,t−1 +b0i gi,t−1 + +γ git = ω0i +γ0i+ + 0i− 1,t−1 1,t−1 k =2
with positivity coefficients, and − βij,t = $0ij + ς0ij+ + 1,t−1 + ς0ij− 1,t−1 +
i X
(k )
τ0ij vk ,t−1 + c0ij βij,t−1
k =2
without positivity constraints. Notice the triangular structure and note that the asymmetry is introduced via the first (observed) factor only.
Dynamic Cholesky decomposition
CHAR models
Estimation
Application
Stationarity for the previous specification There exists a strictly stationary, non anticipative and ergodic solution to the CHAR model when 2 2 − + 1) E log ω01 + γ01+ η1,t−1 + γ01− η1,t−1 + b01 < 0, n o (i) 2) E log α0i ηit2 + b0i < 0 for i = 2, . . . , m, 3) |c0ij | < 1 for all (i, j). Moreover, the stationary solution satisfies Ek1 k2s0 < ∞, Ekg 1 ks0 < ∞, Ekv 1 ks0 < ∞, Ekβ 1 ks0 < ∞ and EkΣ1 ks0 < ∞ for some s0 > 0.
Dynamic Cholesky decomposition
CHAR models
Estimation
Application
Invertibility of the CHAR
Under the stationarity conditions g t = g(η u , u < t),
β t = β(η u , u < t).
For practical use, we need (uniform) invertibility: g t (ϑ) = g(ϑ; u , u < t), with some abuse of notation.
β t (ϑ) = β(ϑ; u , u < t) Invertibility conditions
Dynamic Cholesky decomposition
CHAR models
Estimation
Application
Full QMLE of the general CHAR A QMLE of the CHAR parameter ϑ0 is e n (ϑ), b n = arg min O ϑ
e n (ϑ) = n−1 O
ϑ∈Θ
n X
et (ϑ), q
t=1
e t (ϑ) = Σ (t−1 , . . . , 1 , e where Σ 0 , e −1 , . . . ; ϑ) and 0
−1
e (ϑ)B e (ϑ)G e t (ϑ)t + et (ϑ) = 0t B q t t
m X
eit (ϑ). log g
i=1
• Does not require matrix inversion. • CAN under general regularity conditions.
Regularity conditions
Dynamic Cholesky decomposition
CHAR models
Estimation
Equation-by-Equation (EbE) estimator Consider the triangular model. In a first step, the parameter (1)
ϑ0 = (ω01 , γ01+ , γ01− , b01 ) is estimated by b (1) = arg min ϑ n (1)
ϑ
∈Θ(1)
n X
e1t (ϑ(1) ), q
t=1
where e1t (ϑ(1) ) = q
21t e1t (ϑ(1) ) g
e1t (ϑ(1) ) = ω1 + γ1+ + and g 1t
2
e1t (ϑ(1) ), + log g
+ γ1− − 1t
2
e1,t−1 (ϑ(1) ). + b1 g
Application
Dynamic Cholesky decomposition
CHAR models
Estimation
EbE second step (2) (2) (2) (2) e2t = g e2t (θ 0(2) ). Let ϑ0 = (ϕ0 , θ 0 ), where βe21,t = βe21,t (ϕ0 ) and g (1)
(2)
Independently or in parallel to ϑ0 , one can estimate ϑ0 by b (2) = arg min ϑ n ϑ(2) ∈Θ(2)
n X
e2t (ϑ(2) ), q
t=1
where, for t = 1, . . . , n, ve2t2 (ϕ(2) ) e2t (ϑ(2) ), + log g (2) e g2t (ϑ )
e2t (ϑ(2) ) q
=
e2t (ϑ(2) ) g
(2) 2 e2,t−1 (ϕ(2) ), = ω2,t−1 + α2 ve2,t−1 (ϕ(2) ) + b2 g
ve2t (ϕ(2) )
= 2t − βe21,t (ϕ(2) )1t ,
βe21,t (ϕ(2) )
(2) = ω21,t−1 + τ21 ve2,t−1 (ϕ(2) ) + c21 βe21,t−1 (ϕ(2) ).
Application
Dynamic Cholesky decomposition
CHAR models
Estimation
Application
EbE remaining steps (+i)
For i ≥ 3, βeij,t depends on ϕ0
(i) (−i) (−i) = ϕ0 , ϕ0 , where ϕ0 has
eit depends on been estimated in the previous steps. The volatility g (+i)
ϑ0
(i)
(+i)
= (θ 0 , ϕ0
b (i) = arg min ϑ n
n X
ϑ(i) ∈Θ(i) t=1
(i)
(i)
(i)
), and ϑ0 = (θ 0 , ϕ0 ) can be estimated by
eit (ϑ(i) , ϕ b (−i) q ), n
eit (ϑ(+i) ) = ωi,t−1 + g
i X
eit (ϑ(+i) ) = q
veit2 (ϕ(+i) ) eit (ϑ(+i) ), + log g (+i) e git (ϑ )
(k ) ei,t−1 (ϑ(+i) ), αi vek2,t−1 (ϕ(+k ) ) + bi g
k =2
vekt (ϕ(+k ) ) = kt −
k −1 X
βekj,t (ϕ(+k ) )jt ,
j=1
βeij,t (ϕ(+i) ) = ωij,t−1 +
i X k =2
(k ) τij vek ,t−1 (ϕ(+k ) ) + cij βeij,t−1 (ϕ(+i) ),
Dynamic Cholesky decomposition
CHAR models
Estimation
Application
QML vs. EbE
1) If m = 2, the one-step full QMLE and the two-step EbEE are exactly the same. 2) For m ≥ 3, the two estimators are generally different. 3) The QML and EbE estimators are CAN under similar assumptions.
CAN of the EbEE
4) The EbEE is simpler, but is not always less efficient than the full QMLE.
Example
Similar behaviour on simulations
Dynamic Cholesky decomposition
CHAR models
Estimation
Application
Asset Pricing for Industry Portfolios We consider the 12 industry portfolios used by Engle (2016), examined in the context of the Fama French 3 factor model. The three factors are: MKT (Market factor = excess log-returns of the SP500), SMB (small minus big size factor) and HML (high minus low value factor) Data are from Ken French’s web site and cover the period 1994-2016. We follow Patton and Verardo (2012) in building hedged portfolios to offset some unwanted exposures to predetermined factors.
Dynamic Cholesky decomposition
CHAR models
Estimation
Application
Competing models Let t =
(x0t , yt )0
with xt = (MKTt , SMBt , HMLt )0 and yt = rkt .
Hedging strategy: Et−1 (rkt | x t ) = βk ,MKT ,t MKTt + βk ,SMB,t SMBt + βk ,HML,t HMLt . Competing models: 1) CCC-GARCH(1,1) 2) DCC-GARCH(1,1) 3) CHAR with constant betas 4) CHAR with time varying betas βij,t = $ij + τij vi,t−1 vj,t−1 + cij βij,t−1 CHAR model on (MKTt , SMBt , HMLt , rkt )0 or (MKTt , HMLt , SMBt , rkt )0 by minimizing AIC (or equivalently BIC).
Dynamic Cholesky decomposition
CHAR models
Estimation
Application
BusEq-Mkt 1 2
Buseq: Business Equipment – Computers, Software, and Electronic Equipment
2002
2004
2006
2008
2010
2012
2014
2016
2002
2004
2006
2008
2010
2012
2014
2016
BusEq-SMB -0.5 0 0.5
1
2000
BusEq-HML -2 -1 0
2000
2000
C-CHAR CCC
2002
2004
2006
2008
2010
2012
2014
CHAR DCC
2016
Dynamic Cholesky decomposition
CHAR models
Estimation
Application
Model Confidence Set test (Hansen, Lunde and Nason, 2011) C-CHAR
CHAR
CCC
DCC
BusEq
X
Chems
X
Durbl
X
Enrgy
X
X
Hlth
X
X
Manuf
X
Money
X
NoDur
X
Other
X
Shops
X
Telcm
X
Utils
X
X
Models included in the MCS in the beta hedging exercise. Models highlighted with the symbol X are contained in the model confidence set using a MSE loss function. The significance level for the MCS is set to 20%, and 10,000 bootstrap samples (with a block length of 5 observations).
Dynamic Cholesky decomposition
CHAR models
Estimation
Transaction costs : MKT
SMB
HML
BusEq
0.356
0.380
0.341
Chems
0.310
0.263
0.376
Durbl
0.419
0.464
0.693
Enrgy
0.373
0.337
0.456
Hlth
0.461
0.667
0.397
Manuf
0.442
0.402
0.430
Money
0.390
0.397
0.366
NoDur
0.414
0.383
0.296
Other
0.273
0.343
0.335
Shops
0.344
0.297
0.395
Telcm
0.334
0.414
0.640
Utils
0.465 P1,678
0.408
0.431
∆β
k ,j
=
t=2
∆βCHAR ∆βDCC
|βk ,j,t+1|t − βk ,j,t|t−1 |.
For each column, the figures correspond to the ratio between the value of ∆β
k ,j
obtained for
the CHAR and the DCC-DCB models.
Application
Dynamic Cholesky decomposition
CHAR models
Estimation
Application
Conclusion Compare to other multivariate GARCH (in particular BEKK and DCC), the Cholesky-GARCH models introduced here have several advantages. 1) Precise stationarity and moment conditions exist. 2) The parameters are directly interpretable in terms of DCB. 3) There is no complicated correlation constraint. 4) The estimation can be done without matrix invertion. 5) The asymptotic theory of the QMLE is available. 6) EbE estimation is possible for triangular models. 7) The model works nicely in practice, in particuilar for beta hedging.
Dynamic Cholesky decomposition
CHAR models
Estimation
Application
Conclusion Compare to other multivariate GARCH (in particular BEKK and DCC), the Cholesky-GARCH models introduced here have several advantages. 1) Precise stationarity and moment conditions exist. 2) The parameters are directly interpretable in terms of DCB. 3) There is no complicated correlation constraint. 4) The estimation can be done without matrix invertion. 5) The asymptotic theory of the QMLE is available. 6) EbE estimation is possible for triangular models. 7) The model works nicely in practice, in particuilar for beta hedging. Thanks for your attention!
Dynamic Cholesky decomposition
CHAR models
Estimation
Application
CAN of the EbE estimator Theorem (CAN of the EbEE) bn = Under B1-B5 the EbEE ϑ b n → ϑ0 , ϑ
0
b (1) , . . . , ϑ b (m) ϑ n n
0
0 satisfies
almost surely as n → ∞.
Under the additional assumption B6, as n → ∞, √
(i) L (i) b n ϑn − ϑ0 →N
0, Σ
(i)
:= J
(i)
−1
I
(i)
J
for i = 1 and i = 2. (i)
Jn
=
e (i) (ϑ b (+i) ) ∂2O n n 0 , ∂ϑ(i) ∂ϑ(i)
(i)
In =
n b (+i) ) ∂ e b (+i) ) 1 X ∂e qit (ϑ qit (ϑ n n 0 ∂ϑ(i) ∂ϑ(i)
n t=1
(i)
−1
Dynamic Cholesky decomposition
CHAR models
Estimation
Application
CAN of the EbE estimator (2)
β32,t = $032 + . . . + τ032 v2,t−1 + c032 β32,t−1 , where v2,t−1 = ε2,t−1 − β21,t−1 ε1,t−1 .
Theorem (CAN of the EbEE) (+i)
(+(i−1))
Denoting by Σϕ− (or by Σϕ+ (+i)
) the bottom-right sub-matrix of
(+(i−1))
Σ (or of Σ ) corresponding to the asymptotic variance of b n(−i) (which is equal to ϕ b (+(i−1)) ϕ ), and using the convention n (+2) (2) Σ = Σ , for i = 3, . . . , m we also have √ (+i) (+i) b n ϑ − ϑ0 n
with
Σ
(+i)
=
(i)
Σϑ
0 −1 (i)0 (+(i−1)) − J (i) K Σ + ϕ
(i)
where Σϑ =
L
→N
0, Σ
(+i)
−1 (+(i−1)) − J (i) K (i) Σ + ϕ (+(i−1)) Σ + ϕ
−1 −1 e (i) (ϑ b (+i) ) ∂2 O (i) (+(i−1)) (i)0 n n J (i) and K n = J (i) I (i) + K (i) Σ + K 0 ϕ
∂ϑ(i) ∂ϕ(−i)
Return
Dynamic Cholesky decomposition
CHAR models
Estimation
Invertibility of the CHAR For practical use, we need uniform invertibility: β t (ϑ) = β(ϑ; u , u < t). More precisely, we need
e sup β t (ϑ) − β t (ϑ) ≤ K ρt ,
ϑ∈Θ
e (ϑ) = β(ϑ; t−1 , . . . , 1 , e where β 0 , e −1 , . . . ) for fixed initial t values e 0 , e −1 , . . . , and β t (ϑ0 ) = β t .
Application
Dynamic Cholesky decomposition
CHAR models
Estimation
Application
Invertibility conditions for the "triangular" model In vector form, the model
vekt (ϑ) = kt −
kX −1
βekj,t (ϑ)jt ,
j=1
βeij,t (ϑ) = ωij,t−1 +
i X
(k ) τij vek ,t−1 (ϑ) + cij βeij,t−1 (ϑ),
k =2 − with ωijt = $ij + ςij+ + 1t + ςij− 1t , writes
e t−1 (ϑ)t−1 + C β e (ϑ) =w t−1 + T B e (ϑ) β t t−1 n o 0 e (ϑ) =w t−1 + T t−1 + C − (t−1 ⊗ T )D0m β t−1 e (ϑ). :=w ∗t−1 + S t−1 β t−1
Dynamic Cholesky decomposition
CHAR models
Estimation
Application
Invertibility conditions for the "triangular" model By the Cauchy rule, the triangular model e (ϑ) = w ∗ + S t−1 β e (ϑ). β t t−1 t−1 is uniformly invertible under the conditions E log+ sup kw 1 + T 1 k < ∞, ϑ∈Θ
n
Y
1
γS := lim sup log sup S t−i < 0.
n→∞ n ϑ∈Θ i=1
Return
Dynamic Cholesky decomposition
CHAR models
Estimation
Regularity conditions for the triangular model Let ωit = ωi + γi+ + 1t
2
+ γi− − 1t
2
. The regularity conditions
for CAN of the QMLE of the triangular model
git (ϑ) = ωi,t−1 +
βij,t (ϑ) = ωij,t−1 +
i X k =2 i X
(k )
αi vk2,t−1 (ϑ) + bi gi,t−1 (ϑ) (k )
τij vk ,t−1 (ϑ) + cij βij,t−1 (ϑ)
k =2
are B1 |c0ij | < 1 and other stationarity conditions.
Application
Dynamic Cholesky decomposition
CHAR models
Estimation
Regularity conditions for the triangular model Let ωit = ωi + γi+ + 1t
2
+ γi− − 1t
2
. The regularity conditions
for CAN of the QMLE of the triangular model
git (ϑ) = ωi,t−1 +
βij,t (ϑ) = ωij,t−1 +
i X k =2 i X
(k )
αi vk2,t−1 (ϑ) + bi gi,t−1 (ϑ) (k )
τij vk ,t−1 (ϑ) + cij βij,t−1 (ϑ)
k =2
are B2 For i = 2, . . . , m, the distribution of ηit2 conditionally on {ηjt , j 6= i} is non-degenerate. The support of η1t contains at least two positive points and two negative points.
Application
Dynamic Cholesky decomposition
CHAR models
Estimation
Regularity conditions for the triangular model Let ωit = ωi + γi+ + 1t
2
+ γi− − 1t
2
. The regularity conditions
for CAN of the QMLE of the triangular model
git (ϑ) = ωi,t−1 +
βij,t (ϑ) = ωij,t−1 +
i X k =2 i X
(k )
αi vk2,t−1 (ϑ) + bi gi,t−1 (ϑ) (k )
τij vk ,t−1 (ϑ) + cij βij,t−1 (ϑ)
k =2
are B3 Positivity and invertibility conditions: for all ϑ, (k )
γi+ , γi− , αi
≥ 0 ωi > 0, |bi | < 1 and |cij | < 1.
Application
Dynamic Cholesky decomposition
CHAR models
Estimation
Regularity conditions for the triangular model Let ωit = ωi + γi+ + 1t
2
+ γi− − 1t
2
. The regularity conditions
for CAN of the QMLE of the triangular model
git (ϑ) = ωi,t−1 +
βij,t (ϑ) = ωij,t−1 +
i X k =2 i X
(k )
αi vk2,t−1 (ϑ) + bi gi,t−1 (ϑ) (k )
τij vk ,t−1 (ϑ) + cij βij,t−1 (ϑ)
k =2
are (2) (i) B4 Identifiability conditions: (γ0i+ , γ0i− , α0i , . . . , α0i ) 6= 0 and (2)
(i)
(ς0ij+ , ς0ij− , τ0ij , . . . , τ0ij ) 6= 0.
Application
Dynamic Cholesky decomposition
CHAR models
Estimation
Regularity conditions for the triangular model Let ωit = ωi + γi+ + 1t
2
+ γi− − 1t
2
. The regularity conditions
for CAN of the QMLE of the triangular model
git (ϑ) = ωi,t−1 +
βij,t (ϑ) = ωij,t−1 +
i X k =2 i X
(k )
αi vk2,t−1 (ϑ) + bi gi,t−1 (ϑ) (k )
τij vk ,t−1 (ϑ) + cij βij,t−1 (ϑ)
k =2
are B5 Uniform invertibility condition.
Application
Dynamic Cholesky decomposition
CHAR models
Estimation
Application
Regularity conditions for the triangular model Let ωit = ωi + γi+ + 1t
2
+ γi− − 1t
2
. The regularity conditions
for CAN of the QMLE of the triangular model
git (ϑ) = ωi,t−1 +
βij,t (ϑ) = ωij,t−1 +
i X k =2 i X
(k )
αi vk2,t−1 (ϑ) + bi gi,t−1 (ϑ) (k )
τij vk ,t−1 (ϑ) + cij βij,t−1 (ϑ)
k =2
are B6 Moments conditions (larger than 6) on kt k ,
sup ϑ∈V (ϑ0 )
kβ t (ϑ)k ,
∂β t (ϑ)
sup 0 . ∂ϑ ϑ∈V (ϑ0 )
Return
Dynamic Cholesky decomposition
CHAR models
Estimation
Example Consider a static model with git = 1 for i ∈ {1, 2, 3} and 1t
= v1t
2t
= l21 v1t + v2t
3t
=
l31 v1t + l32 v2t + v3t = l32 v2t + v3t . |{z} 0
In terms of the β’s: 2t
= β21 1t + v2t
3t
= −β21 β32 1t + β32 2t + v3t | {z } β31
→ ϑ = (β21 , β32 ).
Application
Dynamic Cholesky decomposition
CHAR models
Estimation
Application
Example Assume for instance that the variable v2t is independent of the vector (v1t , v3t )0 and that this vector is distributed as the product ηu, where the random variable η and the vector u are p independent, e.g., u ∼ N (0, I 2 ) and η ∼ (ν − 2)/νStν , ν > 4. √ L b QMLE: n ϑn − ϑ0 → N 0, EbEE:
√ n βb21 − β0,21 =
2 Eη 4 1+β32 2 )2 (1+β32
0
P n−1/2 nt=1 η2t 1t L P → n−1 nt=1 21t
. Eη 4 0
N (0, 1) .
Dynamic Cholesky decomposition
CHAR models
Estimation
Application
10
20
(1+β32 2Eη 4)/(1+β32 2)2
30
Illustration
8 7 ν (d
2
egre e
1
6 of fr eedo m
)
5
0 -1
β 32
Figure: Ratio between the asymptotic variance of the QML estimator of β21 and its EbE counterpart as a function of the degree of freedom ν and β32 .
Dynamic Cholesky decomposition
CHAR models
Estimation
Application
Illustration Full QML
EbE
3.5
n × MSE of β^21
3.0
2.5
2.0
1.5
1.0 5.0
7.5
10.0
12.5
15.0
17.5
20.0
22.5
25.0
27.5
30.0
Return
Dynamic Cholesky decomposition
CHAR models
Estimation
Application
Example: Σ = LGL0 , m = 3
1
0
L= l21
1
l31
l32
0 0 , 1
g1
Σ= l21 g1 l31 g1
1
B= −β21 −β31
0 1 −β32
0 0 , 1
l21 g1 2 l21 g1 + g2
l21 l31 g1 + l32 g2
g1 G= 0 0
l31 g1
0
g2
0 ,
0
g3
l21 l31 g1 + l32 g2 . 2 2 l31 g1 + l32 g2 + g3
Remark: In Σ = DRD the constraints on the elements of R are ρ212 + ρ213 + ρ223 − 2ρ12 ρ13 ρ23 ≤ 1. In Σ = LGL0 there is no constraint on the `ij ’s.
0
Dynamic Cholesky decomposition
CHAR models
Estimation
Sequential construction of Σt = Lt G t L0t 1/2
Let the factors v t = G t 1/2
Taking Σt
η t , η t iid (0, I m ) (vit =
1/2
1/2
= Lt G t , we have t = Σt Step 1: v1t = 1t
Step 2:
Step 3:
√
git ηit ).
η t = Lt v t .
g1,t
= Var (v1t | Ft−1 )
2t
= l21,t v1t + v2t
v2t
= 2t − l21,t v1t
g2,t
= Var (v2t | Ft−1 )
3t
= l31,t v1t + l32,t v2t + v3t
v3t
= 3t − l31,t v1t − l32,t v2t
g3,t
= Var (v3t | Ft−1 )
Application
Dynamic Cholesky decomposition
CHAR models
Estimation
Application
= B 0t G −1 Alternative construction of Σ−1 t Bt t t = Lt v t and therefore B t t = v t , where the subdiagonal elements of the lower unitriangular matrix B t = L−1 are −βij,t t Step 1: v1t = 1t
Step 2:
Step 3:
g1,t
= Var (v1t | Ft−1 )
2t
= β21,t 1t + v2t
v2t
= 2t − β21,t 1t
g2,t
= Var (v2t | Ft−1 )
3t
= β31,t 1t + β32,t 2t + v3t
v3t
= 3t − β31,t 1t − β32,t 2t
g3,t
= Var (v3t | Ft−1 ) Return
Dynamic Cholesky decomposition
CHAR models
Estimation
Application
The order of the series Replace t = (1t , 2t , 3t )0 by ∗t = (2t , 1t , 3t )0 (the position of the last component is imposed by the problem) ∗ ∗ ∗ Step 1∗ : v1t = 2t , g1,t = Var (v1t | Ft−1 ) Step 2∗ :
Step 3∗ :
1t
∗ = β12,t 2t + v2t
∗ g2,t
∗ = Var (v2t | Ft−1 )
3t
∗ = β32,t 2t + β31,t 1t + v3t
g3,t
= Var (v3t | Ft−1 )
In particular, β12,t = β21,t β 2
g1t
21,t g1t +g2t
specifications, the order matters.
∗ (v2t 6= v2t )
∗ (v3t = v3t )
: for most parametric
Dynamic Cholesky decomposition
CHAR models
Estimation
Application
The order of the series (matrix form)
Σ−1 t
−1 2 −1 2 g −1 + β21 g2 + β31,t g3,t 1 −1 −1 = −β21 g2 + β31,t β32,t g3,t −1 −β31,t g3,t
∗t = ∆t ,
−1 −β21 g2−1 + β31,t β32,t g3,t −1 2 β32,t g3,t + g2−1 −1 −β32,t g3,t
−1 −β31,t g3,t
−1 −β32,t g3,t . −1 g3,t
∆Σ∗−1 ∆= t
−1 2 β31,t g3,t + g2∗−1
−β12 g ∗−1 + β31,t β32,t g −1 2 3,t −1 −β31 g3,t
−1 −β12 g2∗−1 + β31,t β32,t g3,t −1 2 ∗−1 2 g3,t g2 + β32,t g1∗−1 + β12 −1 −β32,t g3,t
−1 −β31,t g3,t
−1 −β32,t g3,t . −1 g3,t
We have ∆Σ∗−1 ∆ = Σ−1 when t t 2 2 2 2 2 β12 = β21 g1 /(β21 g1 + g2 ), g1∗ = β21 g1 + g2 and g2∗ = g1 − β21 g1 /(β21 g1 + g2 )
When the first parameters are time-invariant, the order does not matter. Return
Dynamic Cholesky decomposition
CHAR models
Estimation
Data Generating process
t
1/2
= Σt
(ϑ0 )η t ,
Σt
= Lt G t L0t ,
gi,t
2 + 0.8gi,t−1 , = 0.1 + 0.1vi,t−1
βij,t
= 0.1 + 0.2vi,t−1 + 0.8βij,t−1 ,
where (η t ) is iid (0, Im ), t = 1, . . . , n.
Application
Dynamic Cholesky decomposition
CHAR models
Estimation
Application
Results for m = 5, η t ∼ N (0, Im ), 1000 replications FULL QML BIAS
RMSE-STD
EbE
5% CP
95% CP
BIAS
RMSE-STD
5% CP
95% CP 92.251
n=1000 ω
0.0202
0.0158
4.444
91.695
0.0190
0.0100
4.070
α
0.0037
0.0024
4.018
91.111
0.0036
-0.0004
3.659
91.696
β
-0.0245
0.0178
7.834
94.007
-0.0230
0.0092
7.379
94.347 92.559
$
0.0008
0.0011
5.769
91.538
0.0007
0.0001
4.604
τ
0.0012
0.0011
7.306
93.939
0.0011
-0.0001
6.208
94.779
c
-0.0017
0.0022
8.115
93.805
-0.0016
0.0003
7.009
94.913
0.0000
0.0050
6.520
92.820
0.0000
0.0021
5.639
93.644 93.138
ALL n=2000 ω
0.0098
0.0014
3.824
92.745
0.0087
0.0000
3.305
α
0.0019
0.0014
4.412
92.966
0.0017
-0.0004
3.766
93.410
β
-0.0119
0.0034
6.642
94.804
-0.0104
-0.0006
6.130
95.460 93.410
$
0.0002
-0.0009
6.483
91.324
0.0002
-0.0003
4.550
τ
0.0003
0.0001
7.145
93.493
0.0004
-0.0002
5.690
94.812
c
-0.0005
-0.0014
8.039
94.069
-0.0004
-0.0006
6.266
95.690
0.0000
0.0002
6.468
93.143
0.0000
-0.0004
5.135
94.426
ALL
2 gi,t = ωi + αi vi,t−1 + βi gi,t−1 and βij,t = $ij + τij vi,t−1 + cij βij,t−1
Dynamic Cholesky decomposition
CHAR models
Estimation
Application
Results for m = 5, η t ∼ t(0, 1, 7), 1000 replications FULL QML BIAS
RMSE-STD
EbE
5% CP
95% CP
BIAS
RMSE-STD
5% CP
95% CP 90.284
n=1000 ω
0.0236
0.0153
4.468
89.811
0.0225
0.0050
4.227
α
0.0056
0.0015
2.931
89.362
0.0057
-0.0066
2.860
90.074
β
-0.0309
0.0181
9.409
92.719
-0.0293
-0.0013
8.980
93.270 92.650
$
0.0010
-0.0002
5.922
91.052
0.0010
-0.0014
4.532
τ
0.0014
-0.0013
7.080
94.173
0.0012
-0.0016
5.910
95.152
c
-0.0019
0.0005
8.818
93.995
-0.0018
-0.0024
7.308
95.226
ALL
-0.0001
0.0037
6.717
92.259
-0.0001
-0.0015
5.730
93.298 91.734
n=2000 ω
0.0062
0.0018
3.624
92.584
0.0093
0.0005
3.141
α
0.0009
0.0007
3.490
91.678
0.0020
-0.0017
2.748
91.189
β
-0.0079
0.0025
7.282
94.899
-0.0120
-0.0006
7.546
94.984 94.079
$
0.0002
0.0001
8.188
90.336
0.0003
-0.0004
4.308
τ
0.0002
0.0002
7.953
92.953
0.0005
-0.0006
4.984
95.507
c
-0.0004
0.0001
9.463
91.695
-0.0007
-0.0008
5.965
95.725
ALL
-0.0001
0.0006
7.289
92.125
0.0000
-0.0006
4.883
94.281
2 gi,t = ωi + αi vi,t−1 + βi gi,t−1 and βij,t = $ij + τij vi,t−1 + cij βij,t−1
Dynamic Cholesky decomposition
CHAR models
Estimation
Application
Results for m = 10, n = 4000 and 1000 replications
Summary statistics on the 165 parameters Mean biais
-0.000873171
Mean (rmse-Mean STD)
0.00155948
Mean Coverage Prob 5%
5.042389091
Mean Coverage Prob 95%
93.6329697 Return
Dynamic Cholesky decomposition
CHAR models
Estimation
Application
The order of the series Replace t = (1t , 2t , 3t )0 by ∗t = (2t , 1t , 3t )0 (the position of the last component is imposed by the problem) ∗ ∗ ∗ Step 1∗ : v1t = 2t , g1,t = Var (v1t | Ft−1 ) Step 2∗ :
Step 3∗ :
1t
∗ = β12,t 2t + v2t
∗ g2,t
∗ = Var (v2t | Ft−1 )
3t
∗ = β32,t 2t + β31,t 1t + v3t
g3,t
= Var (v3t | Ft−1 )
In particular, β12,t = β21,t β 2
g1t
21,t g1t +g2t
specifications, the order matters.
∗ (v2t 6= v2t )
∗ (v3t = v3t )
: for most parametric
Dynamic Cholesky decomposition
CHAR models
Estimation
Application
The order of the series (matrix form)
Σ−1 t
−1 2 −1 2 g −1 + β21 g2 + β31,t g3,t 1 −1 −1 = −β21 g2 + β31,t β32,t g3,t −1 −β31,t g3,t
∗t = ∆t ,
−1 2 β32,t g3,t + g2−1 −1 −β32,t g3,t
−1 −β31,t g3,t
−1 −β32,t g3,t . −1 g3,t
∆Σ∗−1 ∆= t
−1 2 β31,t g3,t + g2∗−1
−1 −β12 g2∗−1 + β31,t β32,t g3,t
−β12 g ∗−1 + β31,t β32,t g −1 2 3,t −1 −β31 g3,t
−1 2 ∗−1 2 g3,t g2 + β32,t g1∗−1 + β12 −1 −β32,t g3,t
We have ∆Σ∗−1 ∆ = Σ−1 when t t β12 =
−1 −β21 g2−1 + β31,t β32,t g3,t
2 β21 g1 /(β21 g1
+ g2 ),
g1∗
=
−1 −β31,t g3,t
−1 −β32,t g3,t . −1 g3,t
Return
2 β21 g1
2 2 2 + g2 and g2∗ = g1 − β21 g1 /(β21 g1 + g2 )
When the first parameters are time-invariant, the order does not matter.