Macroeconomic Forecasting
(MFx)
IMFx
This training material is the property of the International Monetary Fund (IMF) and is intended for use in IMF
Institute for Capacity Development (ICD) courses. Any reuse requires the permission of the ICD.
EViews is a trademark of IHS Global Inc.
Structural Vector Autoregressive Models
(SVAR)
Part 1
Session 1. Introduction to VARs
Introduction
Del Negro and Schorfheide (2011):
At first glance, VARs appear to be
straightforward multivariate generalizations of
univariate autoregressive models. At second
sight, they turn out to be one of the key
empirical tools in modern macroeconomics.
What are VARs ?
multivariate linear time-series models
endogenous variable in the system are functions of lagged
values of all endogenous variables
simple and flexible alternative to the traditional multipleequations models
Historical Overview: Sims Critique
In the 1980s criticized the large-scale
macro-econometric models of the time
Proposed VARs as an alternative that
allowed one to model macroeconomic
data informatively
What Are VARs Used For?
Forecasting
Reduced-Form VARs
Structural Analysis
Structural VARs
Unit Plan/Roadmap
Specification and
Estimation of reducedform VAR
Model
rejected
Model Checking
Forecasting
Source:Lutkepohl (2007)
Model accepted
Structural VAR
specification and
estimation
Impulseresponse
analysis
Forecast-error
variance
decomposition
Session 2. Estimation of VARs
Introduction to VARs
Let yt be a vector with the value of n variables at time t:
y t = [ y1,t y2,t ... yn ,t ]'
A p-order vector autoregressive process generalizes a one-variable AR(p)
process to n variables:
Reduced
y t = G 0 + G 1 y t 1 + G 2 y t-2 + ... + G p y t-p + et
form VAR
G 0 = (n 1) vector of constants
E[et ] = 0
G j = (n n) matrix of coefficients
, if t =
E[et e '] =
0 otherwise
et = (n 1) vector of white noise innovations
NOT diagonal
Example: A VAR(1) in 2 Variables
y1,t = g11 y1,t 1 + g12 y2,t 1 + e1,t
y2,t = g 21 y1,t 1 + g 22 y2,t 1 + e2,t
In matrix notation:
=
y t G 1 y t-1 + e t
Assumptions about the error terms:
where
yt
y1,t
=
, for example: y t
y
2,t
G1
e1,t
g11 g12
=
,
e
t e
g
g
21
22
2,t
gdp
e21
E[et et '] =
e e
12
0
E[et e '] =
0
e e
1 2
e2
2
0
, for t
0
Estimation: by OLS
Performed with OLS applied equation by equation
Estimates are:
consistent
efficient
equivalent to GLS
General Specification Choices
Selection of variables to be included: in accordance with economic
theory, empirical evidence and/or experience
Exogenous variables can be included: constant, time trends, other
additional explanators
Non-stationary level data is often transformed (log levels, log
differences, growth rates, etc.)
The model should be parsimonious
Session 4. Stationary VARs
Stationarity of a VAR: Definition
A p-th order VAR is said to be covariance-stationary if:
1. The expected value of yt does not depend on time
E[ y t ]=
2.
E[ y t + j ]=
1
2
...
n
The covariance matrix of yt and yt+j depends on the time lapsed j and
not on the reference period t
E[( y t )( y t + j ) ']=
E[( y s )( y s + j ) ']=
Conditions for Stationarity
The conditions for a VAR to be stationary are similar to the conditions for a
univariate AR process to be stationary:
y t = G 0 + G 1 y t-1 + G 2 y t-2 + ... + G p y t-p + et
2
(I n G 1L - G 2 L - ... - G p L )y t = G 0 + e t
Lag polynomial
Lag operator
G(L)y t = G 0 + e t
For yt to be stationary, the matrix polynomial in the lag operator G(L) must
be invertible
Conditions for Stationarity
A VAR(p) process is stationary (thus
invertible) if all the np roots of the
characteristic polynomial are (in modulus)
outside the unit imaginary circle
Inverse Roots of AR Characteristic Polynomial
1.5
1.0
0.5
0.0
det(I n G 1L - G 2 L2 - ... - G p Lp ) =
0 -0.5
EViews calculates the inverse roots of the
characteristic AR polynomial, which should
then lie within the unit imaginary circle
-1.0
-1.5
-1.5
-1.0
-0.5
0.0
0.5
1.0
1.5
Vector Moving Average Representation of a VAR
If a VAR is stationary, the yt vector can be expressed as a sum of all of the past
white noise shocks et (VMA() representation)
yt
=
+ G(L)-1 e t , where
y t=
G(L)-1 G 0
+ (I n + 1L + 2 L2 + ...)e t
y t = + et + 1et 1 + 2et 2 + ...
y t = + i et-i
Wold
theorem
i =0
where i is a (n x n) matrix of coefficients, and 0 is the identity matrix.
From the VMA() representation it is possible to obtain impulse response
functions
Session 5. Lag Specification Criteria
Lags Needed for the VAR
What number is most appropriate?
If p is extremely short, the model may be poorly specified
If p is extremely long, too many degrees of freedom will be lost
The number of lags should be sufficient for the residuals from
the estimation to constitute individual white noises
The Curse of Dimensionality
VARs are very densely parametrized
o In a VAR(p) we have p matrices of dimension nxn: G1,.., Gp
o Assume G0 is an intercept vector (dimension: nx1)
o The number of total coefficients/parameters to be estimated is:
n+nxnxp=n(1+nxp)
Overfitting versus Omitted Variable Bias
Over-fitting problem
poor-quality estimates and bad forecasts
Omitted variable bias
Possible solutions:
Core VAR plus rotating variables
Bayesian analysis
Lag Length Criteria
As for univariate models, one can use multidimensional
versions of the:
AIC: Akaike information criterion
SC: Schwarz information criterion
HQ: Hanna-Quinn information criterion
Information-based criteria: trade-off between parsimony
and reduction in sum of squares
Lag Specification: Practitioners Advice
Rules of Thumb:
p = 4 when working with quarterly data
p = 12 with monthly data
The effective constraint is np < T/3
Example:
T = 100
p = 4
n 7
Session 8. Forecasting using VARs
Forecasting Using the Estimated VAR
Let Yt-1 be a matrix containing all information available up to
time t (before realizations of et are known):
Then:
Yt -1 = ( y t -1 , y t -2 ,..., y t -T )
0 +G
1y + G
2 y + ... + G
py
E[y t | Yt-1 ] = G
t-1
t-2
t-p
Forecasting Using the Estimated VAR
The forecast error can be decomposed into the sum of et , the
unexpected innovation of yt , and the coefficient estimation
error:
y t E[ y t | Yt-1 ] =
e t + ( Yt-1 )
If the estimator for the coefficients is consistent and estimates
are based on many data observations, the coefficient
estimation error tends to be small, and:
y t E[ y t | Yt-1 ] e t
Iterated Forecasts
Iterating one period forward:
0 +G
1E[y | Y ] + G
2 y + ... + G
py
E[y t+1 | Yt-1 ] = G
t
t-1
t-1
t-p+1
Iterating j periods forward:
0 +G
1E[y
E[ y t + j | Yt-1 ] = G
t + j-1 | Yt-1 ] + G 2 E[y t + j-2 | Yt-1 ] + ... + G p y t-p+ j
