Beruflich Dokumente
Kultur Dokumente
VECM models
SVAR model
Karolina Konopczak
Karolina Konopczak
VAR model
VECM models
SVAR model
Outline
1
VAR model
Introduction
Specication
Estimation and diagnostics
VECM models
Cointegration
Engle-Granger method / EC model
Johansen procedure / VEC model
SVAR model
SVAR vs. VAR model
Structural analysis
Karolina Konopczak
VAR model
VECM models
SVAR model
Introduction
Specication
Estimation and diagnostics
Plan prezentacji
1
VAR model
Introduction
Specication
Estimation and diagnostics
VECM models
Cointegration
Engle-Granger method / EC model
Johansen procedure / VEC model
SVAR model
SVAR vs. VAR model
Structural analysis
Karolina Konopczak
VAR model
VECM models
SVAR model
Introduction
Specication
Estimation and diagnostics
Introduction (1)
VAR (vector autoregressive) models were developed in the 1980s as an
alternative to multi-equation macroeconomic models (the Cowles
Commission approach) built in the 1950s and 1960s
Sims (1980) critique of the Cowles Commission approach:
ad-hoc specication
arbitrary (theory-based) endogenous/exogenous division of the
model variables
arbitrary (often inadequate) dynamic structure of the model
arbitrary assumptions of zero restrictions in order to pass the
identication conditions
Karolina Konopczak
VAR model
VECM models
SVAR model
Introduction
Specication
Estimation and diagnostics
Introduction (2)
Karolina Konopczak
VAR model
VECM models
SVAR model
Introduction
Specication
Estimation and diagnostics
VAR(p):
xt
m-dimensional
Karolina Konopczak
VAR model
VECM models
SVAR model
Introduction
Specication
Estimation and diagnostics
equations of
Karolina Konopczak
VAR model
VECM models
SVAR model
Introduction
Specication
Estimation and diagnostics
Karolina Konopczak
VAR model
VECM models
SVAR model
Introduction
Specication
Estimation and diagnostics
Canonical representation
VAR(p) can be rewritten as VAR(1)
yt = A1 yt1 + A2 yt2 + ... + Ap ytp + et
yt
A1 A2 . . . Ap1 Ap
yt1 1
0
0
0
= .
..
..
..
..
.
.
.
.
.
.
ytp+1
0
0 ...
1
0
yt1 + ~
yt = A
et > canonical representation
A companion matrix
yt1
yt2
..
.
ytp
of VAR(p)
e.g. VAR(2):
yt = A1 yt1 +
t
A2 yt2 +e
yt
A1 A2
yt1
et
=
+
yt1
1
0
yt2
0
Karolina Konopczak
et
0
..
.
0
VAR model
VECM models
SVAR model
Introduction
Specication
Estimation and diagnostics
operator:
yt = A1 Lyt + A2 L2 yt + ... + Ap Lp yt + et
yt = A(L)yt + et ,
where
A(L) = A1 L + A2 L2 + ... + Ap Lp
(I A(L))yt = et
yt = (I A(L))1 et
yt = (I + A(L) + A(L)2 + ...)et
yt = et + C1 et1 + C2 et2+...
P
yt =
i=0 Ci eti
> yt expressed as a linear combination
error term
Karolina Konopczak
VAR model
VECM models
SVAR model
Introduction
Specication
Estimation and diagnostics
Estimation
since only lagged values of the endogenous variables appear on the
right-hand side of the equation, simultaneity bias is not as issue and OLS
yields consistent estimates
since error terms are contemporaneously correlated, seemingly unrelated
regression (SUR) estimator should be applied in order to yield ecient
estimates
but, since equations have the same set of regressors, SUR estimator
coincides with OLS estimator
in conclusion: parameters can be eciently estimated by OLS for every
equation separately
only if equations do not contain the same exogenous variables (e.g. have
dierent lag structures), SUR should be applied
Karolina Konopczak
VAR model
VECM models
SVAR model
Introduction
Specication
Estimation and diagnostics
we should choose the model with the lowest lag length and residuals
which do not exhibit serial correlation
information
criteria
Karolina Konopczak
VAR model
VECM models
SVAR model
Introduction
Specication
Estimation and diagnostics
Akaike (AIC):
Schwarz
Hannan-Quinn (HQC):
e) +
HQC = ln(det
2K ln(ln(T ))
T
Karolina Konopczak
VAR model
VECM models
SVAR model
Introduction
Specication
Estimation and diagnostics
(1)
(1)
H0
: Ap = 0
(2)
H0
: Ap1 = 0
...
(2)
(p)
vs.
(1)
H1
vs.
: Ap 6= 0
(2)
H1
(p)
: Ap1 6= 0
H0 : A1 = 0 vs. H1 : A1 6= 0
Ap = Ap1 = ... = A2 = 0
(p)
conditionally on
Ap = 0
conditionally on
Karolina Konopczak
VAR model
VECM models
SVAR model
Introduction
Specication
Estimation and diagnostics
Karolina Konopczak
VAR model
VECM models
SVAR model
Introduction
Specication
Estimation and diagnostics
Karolina Konopczak
VAR model
VECM models
SVAR model
Introduction
Specication
Estimation and diagnostics
= C2 = ... = Cq = 0]
T )
no serial correlation up to lag q [E (et eti
Qq = T
where
Pq
j1
= 0 i = 1, ..., q
T )
Cj = E (et etj
Karolina Konopczak
VAR model
VECM models
SVAR model
Introduction
Specication
Estimation and diagnostics
Normality of residuals
Karolina Konopczak
VAR model
VECM models
SVAR model
Introduction
Specication
Estimation and diagnostics
Karolina Konopczak
VAR model
VECM models
SVAR model
Introduction
Specication
Estimation and diagnostics
Granger causality
yft is said to Granger-cause ygt if it contains useful information for
predicting
ygt
y1t
a11i
.
.
.
yft
ygt
..
.
ykt
p
P
af 1i
=
i=1 ag 1i
..
.
ak 1i
...
a1fi
a1gi
...
a1ki
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
...
...
affi
agfi
afgi
aggi
...
...
afki
agki
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
...
akfi
akgi
...
akki
y1ti
e1t
.
.
yfti
ygti
.
ykti
.
.
.
eft
+
egt
..
.
ekt
(RRSSURSS)/m
~
URSS/(T k)
Fm,T k ,
where
Karolina Konopczak
VAR model
VECM models
SVAR model
Cointegration
Engle-Granger method / EC model
Johansen procedure / VEC model
Plan prezentacji
1
VAR model
Introduction
Specication
Estimation and diagnostics
VECM models
Cointegration
Engle-Granger method / EC model
Johansen procedure / VEC model
SVAR model
SVAR vs. VAR model
Structural analysis
Karolina Konopczak
VAR model
VECM models
SVAR model
Cointegration
Engle-Granger method / EC model
Johansen procedure / VEC model
Karolina Konopczak
VAR model
VECM models
SVAR model
Cointegration
Engle-Granger method / EC model
Johansen procedure / VEC model
insignicant
R2
A why?
residuals in spurious regressions are highly correlated / integrated
> variance of residuals is not consistently estimated
> calculated t and F statistics do not follow t and F distributions
Karolina Konopczak
VAR model
VECM models
SVAR model
Cointegration
Engle-Granger method / EC model
Johansen procedure / VEC model
Karolina Konopczak
VAR model
VECM models
SVAR model
Cointegration
Engle-Granger method / EC model
Johansen procedure / VEC model
Cointegration (1)
Denition (2-dimensional case)
2-dimensional vector time series yt is said to be cointegrated of order (d,b) if
both component series, y1t and y2t , are integrated of order d but their linear
combination is integrated of order d-b:
0
yt CI (d, b) y1t I (d) y2t I (d) 6=0 y I (d b)
|{z}
y1 1 +y2 2
= [1 , 2 ] cointegrating vector
Karolina Konopczak
VAR model
VECM models
SVAR model
Cointegration
Engle-Granger method / EC model
Johansen procedure / VEC model
Cointegration (2)
Karolina Konopczak
VAR model
VECM models
SVAR model
Cointegration
Engle-Granger method / EC model
Johansen procedure / VEC model
Cointegration (3)
Karolina Konopczak
VAR model
VECM models
SVAR model
Cointegration
Engle-Granger method / EC model
Johansen procedure / VEC model
Karolina Konopczak
VAR model
VECM models
SVAR model
Cointegration
Engle-Granger method / EC model
Johansen procedure / VEC model
t = t
1 (+
p
P
ti ) + t
i=1
H0 : t I (1) = 0
H1 : t I (0) < 0
CRDF =
S
> the CRDF test is the ADF test performed on estimated residuals (not true values)
> critical values dierent from those from the standard ADF test (MacKinnon, 1991)
Karolina Konopczak
VAR model
VECM models
SVAR model
Cointegration
Engle-Granger method / EC model
Johansen procedure / VEC model
provided that the series are cointegrated, the OLS estimator is:
super-consistent (Stock, 1987)
> estimates converge in probability to the true value at a
faster-than-normal rate
> consistent even if the error term is correlated with the
explanatory variable (no asymptotic simultaneity bias)
not asymptotically normal
> standard errors and hence conventional statistics cannot be used
for statistical inference
often biased due to ignored dynamics (a static model is estimated)
> a serious problem in nite samples
Karolina Konopczak
VAR model
VECM models
SVAR model
Cointegration
Engle-Granger method / EC model
Johansen procedure / VEC model
Karolina Konopczak
VAR model
VECM models
SVAR model
Cointegration
Engle-Granger method / EC model
Johansen procedure / VEC model
yt = 0 + 1 yt1 + 0 xt + 1 xt1 + t
In equilibrium variables will have converged to the steady-state value:
yt = yt1 = y
xt = xt1 = x
Thus:
+1
0
+ 10
x
yt = 1
1
1
y t = 0 + 1 x
Interpretation:
0 short-run multiplier
+1
long-run multiplier
1 = 10
1
Karolina Konopczak
VAR model
VECM models
SVAR model
Cointegration
Engle-Granger method / EC model
Johansen procedure / VEC model
ECM(1,1,1):
+1
0
10
x ) + 0 xt + t
yt = (1 1)(yt1 1
1
1 t1
yt = (yt1 0 1 xt1 ) + 0 xt + t
{z
}
|
ECT : t
1
= 1 1
<0
the larger (in absolute terms), the faster the convergence rate towards
equilibrium
all terms in the ECM are stationary (> standard statistical inference)
Karolina Konopczak
VAR model
VECM models
SVAR model
Cointegration
Engle-Granger method / EC model
Johansen procedure / VEC model
Karolina Konopczak
VAR model
VECM models
SVAR model
Cointegration
Engle-Granger method / EC model
Johansen procedure / VEC model
Karolina Konopczak
VAR model
VECM models
SVAR model
Cointegration
Engle-Granger method / EC model
Johansen procedure / VEC model
yt = yt1 +
Pp1
i=1
= (I
i =
Karolina Konopczak
i yti + t
Pp
i=1
Pp
j=i+1
Ai )
Aj
VAR model
VECM models
SVAR model
Cointegration
Engle-Granger method / EC model
Johansen procedure / VEC model
det(Ik A1 z A2 z 2 ... Ap z p ) = 0
have
characteristic polynomial
det(Ik A1 z A2 z 2 ... Ap z p ) = 0
root (z=1)
Karolina Konopczak
VAR model
VECM models
SVAR model
Cointegration
Engle-Granger method / EC model
Johansen procedure / VEC model
det() = 0
det() = 0
=0
2 possibilities:
(1)
(2) rank(
Karolina Konopczak
VAR model
VECM models
SVAR model
Cointegration
Engle-Granger method / EC model
Johansen procedure / VEC model
Q
rank( ) = number of non-zero eigenvalues = number of cointegrating vectors
Karolina Konopczak
VAR model
VECM models
SVAR model
Cointegration
Engle-Granger method / EC model
Johansen procedure / VEC model
Q
rank( )=k (full rank)
Karolina Konopczak
VAR model
VECM models
SVAR model
Cointegration
Engle-Granger method / EC model
Johansen procedure / VEC model
1 ,
2, ,
k k eigenvalues of the estimated matrix (
) ordered
1 >
2 > >
k
such that
Karolina Konopczak
VAR model
VECM models
SVAR model
Cointegration
Engle-Granger method / EC model
Johansen procedure / VEC model
Trace statistics:
H0 : R = r
H1 : R > r
Pk
TRACE = T i=r +1 ln(1 i )
Maximum eigenvalue statistics:
H0 : R = r
H1 : R = r + 1
r +1 )
MAX = Tln(1
ln(1 i ) = 0 if i = 0,
since
ln(1) = 0
Karolina Konopczak
VAR model
VECM models
SVAR model
Decomposition of the
Cointegration
Engle-Granger method / EC model
Johansen procedure / VEC model
matrix
Q
Q
if rank( )=r<k the
matrix can be decomposed into:
Q
= T
Karolina Konopczak
VAR model
VECM models
SVAR model
Cointegration
Engle-Granger method / EC model
Johansen procedure / VEC model
Hypotheses testing
Karolina Konopczak
VAR model
VECM models
SVAR model
Cointegration
Engle-Granger method / EC model
Johansen procedure / VEC model
Karolina Konopczak
VAR model
VECM models
SVAR model
Plan prezentacji
1
VAR model
Introduction
Specication
Estimation and diagnostics
VECM models
Cointegration
Engle-Granger method / EC model
Johansen procedure / VEC model
SVAR model
SVAR vs. VAR model
Structural analysis
Karolina Konopczak
VAR model
VECM models
SVAR model
Karolina Konopczak
VAR model
VECM models
SVAR model
et = 1 t
Karolina Konopczak
VAR model
VECM models
SVAR model
Structural analysis
Karolina Konopczak
VAR model
VECM models
SVAR model
Identifying restrictions
without additional assumptions (identifying restrictions) structural
parameters cannot be derived (dierent structural models can be
transformed into the same reduced form)
number of restrictions equals the dierence between the number of
parameters of SVAR and of VAR
n(n + 1)
VAR: pn2 +
|{z} | {z
2 }
pAi
n(n + 1)
pn + n pn2
=
|{z} |{z} |{z} | {z
2 }
pB
pA
2
i
n(n1)
Karolina Konopczak
VAR model
VECM models
SVAR model
P= . .
.
.. .. ..
"
#
a q 0
a b
e.g. =
> P = Chol() =
b2
b
c d
d
a
a
Karolina Konopczak
VAR model
VECM models
SVAR model
Karolina Konopczak
VAR model
VECM models
SVAR model
Short-run identication
in general case
we have SVAR: yt = B(L)yt + t , where E (t T
t ) = (
diagonal matrix)
we use triangular (not Cholesky) factorisation
Karolina Konopczak
VAR model
VECM models
SVAR model
Karolina Konopczak
VAR model
VECM models
SVAR model
Ci eti , where et = 1 t
P
P
d11i
SVARMA: yt =
=
i=
0
i=0 Di ti
|{z}
d21i
VARMA: yt =
i=0
where Di = 1 Ci
IRFknj =
yk
yn
tj
bivariate model
d12i
d22i
1
yti
2
yti
t1
y2
y1 ,
t
1
d221 =
y2
y2
t
1
since endogenous variables are stationary structural shocks fade away, i.e.
have no long-run impact on the level of yk :
lim dknj = 0
j
Karolina Konopczak
VAR model
VECM models
SVAR model
Cumulative (long-run)
impact of structural
shocks:
P
P
P
d
d
d11
11
i
12
i
i=
0
i=
0
P
D(1) = i=0 Di = P
=
d31
i=0 d21i
i=0 d22i
Karolina Konopczak
d12
d32
VAR model
VECM models
SVAR model
Derivation (1):
SVARMA representation for yt+j :
yt+j = D0 t+j + D1 t+j1 + ... + Dj1 t+1 + Dj t + Dj+1 t1 + ...
the best linear forecast of yt+j based on information available at time t:
Eyt+j|t = Dj t + Dj+1 t1 + ...
forecast error:
P 1
yt+j Eyt+j|t = D0 t+j + D1 t+j1 + ... + Dj1 t+1 = j
i=0 Di t+ji
Karolina Konopczak
VAR model
VECM models
SVAR model
FEVD > the relative share of variance that n-th structural shock
contributes to the
forecast error variance of k-th variable:
Pjtotal
1 d2
2
FEVDknj = n i=20 kn,i
kj
Karolina Konopczak