THÈSE Estimation, validation et identification des modèles ARMA ...

Chapitre 5. Model selection of weak VARMA models 160
with equality if and only if e1(θ) = e1(θ0) a.s. Using the elementary inequality
Tr(A −1 B)−logdet(A −1 B) ≥ Tr(A −1 A)−logdet(A −1 A) = d for all symmetric positive
semi-definite matrices of order d×d, it is easy see that ∆(θ)−∆(θ0) ≥ 0. The proof is
complete. ✷
5.5.1 Estimating the discrepancy
Let X = (X1,...,Xn) be observation of a process satisfying the VARMA representation
(5.1). Let ˆ θn be the QMLE of the parameter θ of a candidate VARMA model.
Let, êt = ˜et( ˆ θn) be the QMLE/LSE residuals when p > 0 or q > 0, and let êt = et = Xt
when p = q = 0. When p+q = 0, we have êt = 0 for t ≤ 0 and t > n, and
êt = Xt −
p
i=1
A −1
0 (ˆ θn)Ai( ˆ θn) ˆ Xt−i +
for t = 1,...,n, with ˆ Xt = 0 for t ≤ 0 and ˆ Xt = Xt for t ≥ 1.
q
i=1
A −1
0 (ˆ θn)Bi( ˆ θn)B −1
0 (ˆ θn)A0( ˆ θn)êt−i,
In view of Lemma 5.1, it is natural to minimize an estimation of the theoretical
criterion E∆( ˆ θn). Note that E∆( ˆ θn) can be interpreted as the average discrepancy
when one uses the model of parameter ˆ θn. The Akaike information criterion (AIC)
is an approximately unbiased estimator of E∆( ˆ θn). We will adapt to weak VARMA
models the corrected AIC version (AICc) introduced by Tsai and Hurvich (1989) for the
univariate strong AR models. Under Assumptions A1–A9, an approximately unbiased
estimator of E∆( ˆ θn) is given by
AICM = nlogdet ˆ Σe + n2d2 nd
+ vec
nd−k1 2(nd−k1)
Î′ ′
11,n vec ˆ J −1
11,n , (5.4)
with vec ˆ J11,n and vec Î11,n are defined in Section 5.2. The AICM stands for AIC "modified".
Justification of (5.4). Using a Taylor expansion of the functions ∂logLn( ˆ θn)/∂θ (1)
, it follows that
around θ (1)
0
ˆθ (1)
n −θ(1)
1
op √n
0 = − 2
n J−1 11
∂logLn(θ0)
∂θ (1)
= − 2
n
n
t=1
J −1
11
where a c = b signifies a = b+c and where J11 = J11(θ0) with
2∂
J11(θ) = lim
n→∞ n
2logLn(θ) ∂θ (1) ∂θ (1)′ a.s.
∂e ′ t(θ0)
Σ−1
∂θ (1) e0 et(θ0), (5.5)