Professional Documents
Culture Documents
) =
_
_
_
t =
0 t =
positive denite matrix.
Eduardo Rossi c - Time Series Econometrics 11 2
VAR(p)
A VAR is a vector generalization of a scalar autoregression.
The VAR is a system in which each variable is regressed on a
constant and p of its own lags as well as on p lags of each of the other
variables in the VAR.
[I
N
1
L
p
L
p
] y
t
= c +
t
(L)y
t
= c +
t
with
(L) = [I
N
1
L
p
L
p
]
(L) (N N) matrix polynomial in L.
Eduardo Rossi c - Time Series Econometrics 11 3
VAR(p) - Stationarity
The element (i, j) in (L) is a scalar polynomial in L
ij
(1)
ij
L
(2)
ij
L
2
. . .
(p)
ij
L
p
ij
=
_
_
_
1 i = j
0 i = j
Stationarity: A vector process is said covariance stationary if its
rst and second moments, E[y
t
] and E[y
t
y
tj
] respectively, are
independent of the date t.
Eduardo Rossi c - Time Series Econometrics 11 4
VAR(1)
y
t
= c +
1
y
t1
+
t
First equation:
y
1t
= c
1
+
(1)
11
y
1t1
+
(1)
12
y
2t1
+ . . . +
(1)
1N
y
Nt1
+
1t
y
t
= c +
1
[c +
1
y
t2
+
t1
] +
t
= c +
1
c +
2
1
y
t2
+
t
+
1
t1
y
t
= . . .
y
t
= c +
1
c + . . . +
k1
1
c +
k
1
y
tk
+
t
+
1
t1
+ . . . +
k1
1
tk+1
E[y
t
] =
k1
j=0
j
1
c +
k
1
E[y
tk
]
The value of this sum depends on the behavior of
j
1
as j increases.
Eduardo Rossi c - Time Series Econometrics 11 5
Stability of VAR(1)
Let
1
,
2
, . . . ,
N
be the eigenvalues of
1
, the solutions to the
characteristic equation
|
1
I
N
| = 0
then, if the eigenvalues are all distinct
1
= QMQ
1
k
1
= QM
k
Q
1
M
k
= diag(
k
1
,
k
2
, . . . ,
k
N
)
If |
i
| < 1, i = 1, . . . , N
k
0, k
If |
i
| 1, i then one or more elements of M
k
are not vanishing,
and may be tending to .
Eduardo Rossi c - Time Series Econometrics 11 6
VAR(P) Companion form
VAR(p):
y
t
=
1
y
t1
+ . . . +
p
y
tp
+
t
as a VAR(1) (Companion Form):
t
= F
t1
+v
t
t
=
_
_
y
t
.
.
.
y
tp+1
_
_
(Np 1)
t1
=
_
_
y
t1
.
.
.
y
tp
_
_
(Np 1)
Eduardo Rossi c - Time Series Econometrics 11 7
VAR(P) Companion form
F =
_
1
2
. . .
p1
p
I
N
0 . . . 0 0
0 I
N
. . . 0 0
.
.
.
.
.
.
.
.
.
.
.
.
0 0 . . . I
N
0
_
_
(NpNp) v
t
=
_
t
0
.
.
.
0
_
_
(Np1)
E[v
t
v
] =
_
_
_
Q t =
0 t =
Q =
_
_
0 . . . 0
0 0 . . . 0
.
.
.
.
.
.
.
.
.
0 0 . . . 0
_
_
(Np Np)
Eduardo Rossi c - Time Series Econometrics 11 8
Stability of VAR(p)
y
t
=
1
y
t1
+ . . . +
p
y
tp
+
t
t = 1, . . . , T
If the process y
t
has nite variance and an autocovariance sequence
that converges to zero at an exponential rate, then
t
must share
these properties. This is ensured by having the Np eigenvalues of F
lie inside the unit circle.
The determinant dening the characteristic equation is
|F I
Np
| = (1)
Np
|
p
I
N
p1
1
. . .
p
| = 0
Eduardo Rossi c - Time Series Econometrics 11 9
Stability of VAR(p)
The required condition is that the roots of the equation
|
p
I
N
p1
1
. . .
p
| = 0
a polynomial of order Np must lie inside the unit circle.
Stability condition can also be expressed as the roots of
|(z)| = 0
lie outside the unit circle, where (z) is a (N N) matrix
polynomial in the lag operator of order p.
Eduardo Rossi c - Time Series Econometrics 11 10
Stability of VAR(p)
When p = 1, the roots of
|I
N
1
z| = 0
outside the unit circle, i.e. |z| > 1, implies that the eigenvalues of
1
be inside the unit circle. Note that the eigenvalues, roots of
|
1
I
N
| = 0, are the reciprocal of the roots of |I
N
1
z| = 0.
Eduardo Rossi c - Time Series Econometrics 11 11
Stability of VAR(p)
Three conditions are necessary for stationarity of the VAR(p) model:
Absence of mean shifts;
The vectors {
t
} are identically distributed, t;
Stability condition on F.
If the process is covariance stationary we can take the expectations of
both sides of
y
t
= c +
1
y
t1
+ . . . +
p
y
tp
+
t
= c +
1
+ . . . +
p
= [I
N
1
. . .
p
]
1
c
= (1)
1
c
Eduardo Rossi c - Time Series Econometrics 11 12
Vector MA()
If the VAR(p) is stationary then it has a VMA() representation:
y
t
= +
t
+
1
t1
+
2
t2
+ . . . +(L)
t
y
tj
is a linear function of
tj
,
tj1
, . . . each of which is
uncorrelated with
t+1
for j = 0, 1, . . ..
It follows that
t+1
is uncorrelated with y
tj
for any j 0.
Linear forecast of y
t+1
based on y
t
, y
t1
, . . . is given by
y
t+1|t
= +
1
(y
t
) +
2
(y
t1
) +. . . +(y
tp+1
)
Eduardo Rossi c - Time Series Econometrics 11 13
Forecasting with VAR
t+1
can be interpreted as the fundamental innovation in y
t+1
, that
is the error in forecasting y
t+1
on the basis of a linear function of a
constant and y
t
, y
t1
, . . ..
A forecast of y
t+s
on the basis of y
t
, y
t1
, . . . will take the form
y
t+s|t
= +F
(s)
11
(y
t
) +F
(s)
12
(y
t1
) +. . . +F
(s)
1p
(y
tp+1
)
The moving average matrices
j
can be calculated as:
(L) = [(L)]
1
(L)(L) = I
N
Eduardo Rossi c - Time Series Econometrics 11 14
VMA coefficient matrices
[I
N
+
1
L +
2
L
2
+ . . .][I
N
1
L
2
L
2
+ . . .
p
L
p
] = I
N
Setting the coecient on L
1
equal to the zero matrix,
1
= 0
on L
2
2
=
1
1
+
2
In general for L
s
s
=
1
s1
+ . . . +
p
sp
s = 1, 2, . . .
0
= I
N
,
s
= 0 s < 0
Eduardo Rossi c - Time Series Econometrics 11 15
VMA coefficient matrices
The innovation in the MA() representation is
t
, the fundamental
innovation for y.
There are alternative MA representation based on VWN processes
other than
t
.
Let H be a nonsingular (N N) matrix
u
t
H
t
u
t
V WN.
y
t
= +H
1
H
t
+
1
H
1
H
t1
+ . . .
= +J
0
u
t
+J
1
u
t1
+J
2
u
t2
+ . . .
J
s
s
H
1
Eduardo Rossi c - Time Series Econometrics 11 16
VMA coefficient matrices
For example H can be any matrix that diagonalizes , the var-cov of
t
,
HH
= D
the elements of u
t
are uncorrelated with one another.
It is always possible to write a stationary VAR(p) process as a
convergent innite MA of a VWN whose elements are mutually
uncorrelated.
To obtain the MA representation for the fundamental innovations, we
must impose
0
= I
N
(while J
0
is not the identity matrix).
Eduardo Rossi c - Time Series Econometrics 11 17
Assumptions implicit in a VAR
For a covariance stationary process, the parameters c,
1
, . . . ,
p
could be dened as the coecients of the projections of y
t
on
1, y
t1
, y
t2
, . . . , y
tp
.
t
is uncorrelated with y
t1
, . . . , y
tp
by the denition of
1
, . . . ,
p
.
The parameters of a VAR can be estimated consistently with n
OLS regressions.
t
dened by this projection is uncorrelated with
y
tp1
, y
tp2
, . . .
The assumption of y
t
V AR(p) is basically the assumption that
p lags are sucient to summarize the dynamic correlations
between elements of y.
Eduardo Rossi c - Time Series Econometrics 11 18
Vector MA(q) Process
y
t
= +
t
+
1
t1
+ . . . +
q
tq
t
V WN,
j
(N N) matrix of MA coecients j = 1, 2, . . . , q.
E(y
t
) =
0
= E[(y
t
)(y
t
)
]
= E[
t
t
] +
1
E[
t1
t1
]
1
+ . . . +
q
E[
tq
tq
]
q
= +
1
1
+ . . . +
q
q
Eduardo Rossi c - Time Series Econometrics 11 19
Vector MA(q) Process
j
= E[(
t
+
1
t1
+ . . . +
q
tq
)(
tj
+
1
tj1
+ . . . +
q
tjq
)
j
=
_
j
+
j+1
1
+
j+2
2
+ . . . +
q
q1
j = 1, . . . , q
j
+
1
j+1
+ . . . +
q+j
q
j = 1, . . . , q
0 |j| > q
0
= I
N
. Any VMA() is covariance stationary.
Eduardo Rossi c - Time Series Econometrics 11 20
VAR(p) Autocovariances
Given:
j
= E[(y
t
)(y
tj
)
j
= E[(y
t
)(y
t+j
)
]
then
j
=
j
j
=
j
{
j
}
1,2
= cov(y
1t
, y
2tj
)
{
j
}
2,1
= cov(y
2t
, y
1tj
)
{
j
}
1,2
= cov(y
1t
, y
2t+j
)
j
= E[(y
t+j
)(y
(t+j)j
)
]
= E[(y
t+j
)(y
t
)
j
= E[(y
t
)(y
t+j
)
] =
j
Eduardo Rossi c - Time Series Econometrics 11 21
VAR(p) Autocovariances
Companion form:
t
= F
t1
+v
t
= E[
t
t
]
= E
_
_
_
_
y
t
y
t1
.
.
.
y
tp+1
_
[(y
t
)
. . . (y
tp+1
)
]
_
_
=
_
0
1
. . .
p1
1
0
. . .
p2
.
.
.
.
.
.
p1
p2
. . .
0
_
_
(Np Np)
Eduardo Rossi c - Time Series Econometrics 11 22
VAR(p) Autocovariances
E[
t
t
] = E[(F
t1
+v
t
)(F
t1
+v
t
)
]
= FE(
t1
t1
)F
+ E(v
t
v
t
)
where FE(
t1
v
t
) = 0.
= FF
+Q
vec() = vec(FF
) + vec(Q)
vec() = vec(FF
) + vec(Q)
vec() = (F F)vec() + vec(Q)
= [I
(Np)
2 (F F)]
1
vec(Q)
The eigenvalues of (F F) are all of the form
j
j
where
i
and
j
are eigenvalues of F. Since |
i
| < 1, i, it follows that all eigenvalues
of (F F) are inside the unit circle |I
(Np)
2 (F F)| = 0.
Eduardo Rossi c - Time Series Econometrics 11 23
VAR(p) Autocovariances
The j -th autocovariance of
t
is
E[
t
tj
] = FE[
t1
tj
] + E[v
t
tj
]
j
= F
j1
j
= F
j
j = 1, 2, . . .
The j -th autocovariance of y
t
is given by the rst rows and n
columns of
j
j
=
1
j1
+ . . . +
p
jp
Eduardo Rossi c - Time Series Econometrics 11 24
Maximum Likelihood Estimation
y
t
= c +
1
y
t1
+ . . . +
p
y
tp
+
t
(1)
t
i.i.d.N(0, )
(T + p) observations. Conditioning on the rst p observations we
estimate using the last T observations.
Conditional likelihood:
f
Y
T
,Y
T1
,...,Y
1
|Y
0
,...,Y
1p
(y
T
, . . . , y
1
|y
0
, . . . , y
1p
; )
Eduardo Rossi c - Time Series Econometrics 11 25
Maximum Likelihood Estimation
= (c
, vec(
1
)
, vec(
2
)
, . . . , vec(
p
)
, vech()
y
t
|y
t1
, . . . , y
p+1
N(c +
1
y
t1
+ . . . +
p
y
tp
, )
x
t
=
_
_
1
y
t1
.
.
.
y
tp
_
_
(Np + 1) 1
_
c
1
2
. . .
p
_
(N (Np + 1))
E(y
t
|y
t1
, . . . , y
p+1
) =
x
t
Eduardo Rossi c - Time Series Econometrics 11 26
Maximum Likelihood Estimation
f
Y
t
|Y
t1
,...,Y
1p
(y
t
|y
t1
, . . . , y
1p
; ) =
(2)
n
2
|
1
|
1
2
exp
_
1
2
(y
t
x
t
)
1
(y
t
x
t
)
_
The joint density, conditional on the y
0
, . . . , y
1p
f
Y
t
,...,Y
1
|Y
0
,...,Y
1p
(y
t
, . . . , y
1
|y
0
, . . . , y
1p
; ) = f
Y
t
|Y
t1
,...,Y
1p
()
f
Y
t1
,...,Y
1
|Y
0
,...,Y
1p
()
Eduardo Rossi c - Time Series Econometrics 11 27
Maximum Likelihood Estimation
Recursively, the likelihood function for the full sample, conditioning
on y
0
, . . . , y
1p
, is the product of the single conditional densities:
f
Y
T
,...,Y
1
|Y
0
,...,Y
1p
=
T
t=1
f
Y
t
|Y
t1
,...,Y
1p
(y
t
|y
t1
, . . . , y
1p
; )
The log likelihood function:
L() =
TN
2
log (2)+
T
2
log |
1
|
1
2
T
t=1
_
(y
t
x
t
)
1
(y
t
x
t
)
=
_
T
t=1
y
t
x
t
__
T
t
x
t
x
t=1
_
1
The j -th row of
is:
u
= u
j
_
T
t=1
y
t
x
t
__
T
t
x
t
x
t
_
1
This is the estimated coecient vector from an OLS regression of y
jt
on x
t
. ML estimates are found by an OLS regression of y
jt
on a
constant and p lags of all the variables in the system.
The ML estimate of is given by:
=
1
T
T
t=1
t
Eduardo Rossi c - Time Series Econometrics 11 29
Maximum Likelihood Estimation
Asymptotic distribution of
The ML estimates
and
will give consistent estimates of the
population parameters even if the true innovations are non-gaussian:
(L)y
t
=
t
t
i.i.d.(0, )
E[
it
jt
lt
mt
] < i, j, l, m
the roots of
|I
N
1
z . . .
p
z
p
| = 0
Let K N p + 1, x
t
(1 K):
x
t
= [1 , y
t
, y
t1
, . . . , y
tp
]
T
= vec(
T
)
Eduardo Rossi c - Time Series Econometrics 11 30
Maximum Likelihood Estimation
T
=
_
_
1,T
.
.
.
N,T
_
_
i,t
=
_
t
x
t
x
t
_
1
_
t
x
t
y
it
_
T
=
1
T
it
= y
it
x
t
i,T
Eduardo Rossi c - Time Series Econometrics 11 31
Maximum Likelihood Estimation
Then,
1
T
t
x
t
x
t
p
Q = E (x
t
x
t
)
T
p
T
p
T(
T
)
d
N(0, (Q
1
))
T(
i,T
)
d
N(0, (
2
i
Q
1
))
2
i
= E(
2
it
)
2
i
=
1
T
2
it
p
2
i
Eduardo Rossi c - Time Series Econometrics 11 32
Maximum Likelihood Estimation
i
N
_
_
i
,
2
i
_
t
x
t
x
t
_
1
_
_
OLS t and F statistics applied to the coecients of any single
equation in the VAR are asymptotically valid. A more general
hypothesis
R = d
can be tested using a generalization of the Wald form of the OLS
2
test
T(R
T
d)
d
N
_
0, R
_
Q
1
_
R
_
Eduardo Rossi c - Time Series Econometrics 11 33
Maximum Likelihood Estimation
2
(m) = T(R
T
d)
_
R
_
T
Q
1
T
_
R
_
1
(R
T
d)
= (R
T
d)
_
R
_
T
(T
Q
T
)
1
_
R
_
1
(R
T
d)
= (R
T
d)
_
_
R
_
_
T
_
t
x
t
x
t
_
1
_
_
R
_
_
1
(R
T
d)
Eduardo Rossi c - Time Series Econometrics 11 34