See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/315779082

VAR, VEC and Toda-Yamamoto Models

Presentation · November 2015

CITATIONS READS

0 858

1 author:

Afees Salisu
University of Ibadan
89 PUBLICATIONS 291 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Rethinking the Phillips curve View project

IIPELP Econometrics View project

All content following this page was uploaded by Afees Salisu on 04 April 2017.

The user has requested enhancement of the downloaded file.

ECONOMETRICS FOR RESEARCHERS, 2015
(MODULE IV: November 16-20, 2015)

Multivariate Models

Afees A. Salisu

Contacts: Department of Economics, Federal University of Agriculture,

Abeokuta & Centre for Econometric & Allied Research (CEAR), University

of Ibadan. Email: adebare1@yahoo.com; aa.salisu@cear.org.ng.

Mobile: +234 (0)803 471 1769.

1
 Multivariate Models
Introduction
The vector autoregression (VAR) model is one of the most successful, flexible, and easy
to use models for the analysis of multivariate time series. It is a natural extension of the
univariate autoregressive model to dynamic multivariate time series. The VAR model
has proven to be especially useful for describing the dynamic behavior of economic and
financial time series and for forecasting. It often provides superior forecasts to those
from univariate time series models. Forecasts from VAR models are quite flexible
because they can be made conditional on the potential future paths of specified
variables in the model.

Unlike the simultaneous, or structural, equation models where some variables are
treated as endogenous and some as exogenous or predetermined (exogenous plus
lagged endogenous), in VAR models all the variables are treated as endogenous and
therefore, there is no a priori distinction between endogenous and exogenous variables.

In addition to data description and forecasting, the VAR model is also used for
structural inference and policy analysis. In structural analysis, certain assumptions
about the causal structure of the data under investigation are imposed, and the
resulting causal impacts of unexpected shocks or innovations to specified variables on
the variables in the model are summarized. These causal impacts are usually
summarized with impulse response functions and forecast error variance
decompositions.

There are now different extensions of VAR developed to capture the underlying
statistical features of the series under examination. Among the prominent extensions
are the Vector Error Correction, VAR-in-Differences and Toda-Yamamoto Models.

Therefore, at the end of this session, participants should be able to:

 Specify and estimate multivariate models,

 Perform relevant diagnostic tests,
 Carry out causality tests,
 Compute impulse response functions
 Compute forecast error variance decompositions
 Interpret Results

2
Estimation Procedure for VAR/VECM/TODA-YAMAMOTO
The estimation procedure can be partitioned into three to include:
I. Data and Pre-Estimation/Preliminary Analysis
 Data Issues
 Graphical Analysis
 Descriptive statistics
 Formal Pre-tests (Unit root and Cointegration tests)

II. Estimation
 VAR
 VECM
 Toda-Yamamoto
 Impulse Response Function (IRF)
 Forecast Error Variance Decomposition (FEVD)
 Granger Non-causality

III. Post Estimation

 Serial Correlation
 Heteroscedasticity
 Normality

3
 Research Paper Title:
The dynamic relationship between stock and money markets in
Nigeria

I. Data and Preliminary Analysis

1.0 Data
Essentially, this study covers two variables namely the stock market and the money
market with the former proxied by All Share Index (ASI) of the Nigerian Stock
Exchange while the latter is captured by the government bond yield (i.e. Treasury
Bills). The ASI covers all the listed equities on the Exchange, including those listed
on the Alternative Securities Market (ASeM), regardless of capitalization. The
treasury bills (T-bills) used here is the short term (3-month) T-bills. The long term (6-
month) T-bills is not considered due to data paucity; the publication of the data only
commenced in 2008. The variables are sourced from the Central Bank of Nigeria’s
Statistical bulletin over the monthly period of January 2000 to September 2015.

2.0 Graphical Representation

 The graphical representation of the two series in level form is shown in Figure 1.
 The combined graphical representation of stock market (ASI) and money market
(3M_TBILLS) in the figure below shows that the two markets are moving in
opposite directions.

4
Figure 1: Relationship between Stock Market and Money Market in Nigeria, 2000:M1-2015:M7
80,000

60,000

40,000

40
20,000

30
0

20

10

0
00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15

ASI 3M_TBILLS

3.0 Descriptive Statistics

 Table 1 below summarizes the basic statistical features of the data under
consideration including the mean, the minimum and maximum values, standard
deviation, skewness, kurtosis and the Jarque-Bera test for the data.
 These descriptive statistics provide a historical background for the behavior of
our data.
 There seems to be evidence of significant variations as shown by the huge
difference between the minimum and maximum values for the variables under
consideration.

Table 1: Descriptive Statistics

Log(ASI) Log(TBR)
Mean 10.110 2.323
Median 10.097 2.379
Maximum 11.092 3.449
Minimum 9.082 0.039
Std. Dev. 0.449 0.646
Skewness -0.217 -0.846
Kurtosis 2.777 4.055
Jarque-Bera 1.852 (0.396) 30.978 (0.000)
Observation 187 187

5
  The skewness of the data series indicates an asymmetric or non-normal data
distribution as the series relatively deviates from normality maintaining negative
skewness.
 The kurtosis statistic equally shows that Log(ASI) is platykurtic in nature while
Log(TBR) on the other hand is leptokurtic.
 The Jarque-Bera test is a test of normality. The null hypothesis for the test is that
the series under consideration is normally distributed.
 Based on our results using the P-values associated with the Jarque-Bera statistics,
the log(ASI) is normally distributed while the log(TBR) is not.
 Note that the multivariate framework does not require the normality assumption
 However, the likelihood ratio function in the Johansen Cointegration test relies
on the assumption.

4.0 The Formal Pre-Tests

4.1 Unit Root Test

 The unit root tests considered include the conventional unit root tests; namely
Augmented Dickey-Fuller (ADF), Phillips-Perron (PP), Dickey-Fuller-GLS (DF-
GLS), Ng-Perron (Ng-P), Kwiatkwoski-Phillips-Schmidt-Shin (KPSS) while
Perron (2005) is unit root test with structural breaks.
 The null hypothesis for ADF, PP, DF-GLS and Ng-P is that an observable time
series is not stationary (i.e. has unit root) while that of KPSS tests for the null
hypothesis is that the series is trend stationary.
 The Perron (2005) test accounts for structural breaks while testing for unit root,
The idea is to confirm that the unit root observed for a particular series is not due
to structural breaks.
 The critical values can be obtained from Table 1(e) for Model 2 in Perron (2005).
They are -5.28 and -4.62 respectively for 1% and 5% levels of statistical
significance.

6
  Panel A in Table 2 reports the unit root test results for the series in their level and
difference forms considering all the available test options except where not
applicable.
 Test options considered are constant (a), constant and trend (b) and without
constant and trend (n).
 Panel B, Table 2 also reports the results of unit root with structural break in levels
and first differences.
 Table 3 is the summary of the unit root test results.
Table 2: Unit Root and Stationarity Tests
Panel A: Unit Root without Structural Break
Variable: Log(ASI)
Level Difference
Constant Constant & None Constant Constant & None
Trend Trend
ADF -2.266 -1.750 1.127 -12.008* -12.089* -11.962*
PP -2.287 -1.933 0.933 -11.997* -12.081* -12.008*
DF- -2.577 -1.008 Not applicable -6.948* Not applicable
GLS -11.782*
Ng-P -0.024 -0.984 Not applicable -5.433* -6.719* Not applicable
KPSS 0.697* 0.248 Not applicable 0.225 * 0.084 * Not applicable

Variable: Log(TBR)
ADF -2.744*** -2.826 -1.003 -10.826* -10.815* -10.849*
PP -2.435 -2.209 -0.975 -10.757* -10.675* -10.782*
DFGLS -1.610 -2.665*** Not applicable -10.819* -10.838* Not applicable
Ng-P -1.6142 -2.665*** Not applicable -6.631* -6.633* Not applicable
KPSS 0.699* 0.323 Not applicable 0.061* 0.025* Not applicable

Panel B: Unit Root with Structural Break: Perron (2005)

Level First Difference
Variable Break Date Coefficient T. stat Break Date Coefficient T. stat
Log(ASI) 2007:08 -0.072 -3.883 2008:05 -1.136 -5.501
Log(TBR) 2010:07 -0.203 -5.088 2003:12 -0.956 -5.551

Note: *, ** and *** imply significance at 1%, 5% and 10% respectively.

7
 Table 3: Summary of Unit Root Results
3A: Unit Root without Structural Breaks
Log(ASI) Log(TBR)
Level First Diff. I(d) Level First Diff. I(d)
ADF -2.266a -12.089 b* I(1) -2.744 a*** I(0)
PP -2.287a -12.081 b* I(1) -2.435 -10.782*n I(1)
DF-GLS -2.577a -11.782b* I(1) -2.665b* I(0)
Ng-P -0.984b -6.719b* I(1) -2.665b* I(0)

3B: Unit Root with Structural Breaks

Break Date Level Fist Diff. I(d)
Log(ASI) 2008:05 -3.883 -5.501* I(1)
Log(TBR) 2010:07 -5.088** - I(0)
Note: *, ** and *** imply statistical significance at 1%, 5% and 10% levels
respectively. Also, a denotes model with constant, b model with constant and trend
and n is model without constant and trend.

4.2 Cointegration Test

 There are alternative co-integration tests in the literature. The suitability of a
particular approach however depends on the underlying test equation and the
order of integration of the series under consideration. Essentially, there are two
test equations for cointegration analyses. We have the single-equation and the
multiple-equation cointegration tests. The former is meant for single-equation
models such as the simple and multiple regression models while the latter is
specifically designed for multivariate models. The prominent examples for the
single-equation case are the Engle-Granger cointegration test and the Bounds
Cointegration test while for the multiple case; we have the Johansen
Cointegration test although some researchers tend to use the latter for a single-
equation case. However, it is not acceptable to use the single-equation test for
multivariate models.
 The second condition which has to do with the order of integration relates to the
stationarity properties of the underlying variables.

8
  The Engle-Granger cointegration test and the Johansen cointegration test require
that all the series must be integrated of order 1 (i.e. I(1)) while the Bounds
cointegration test allows for the combination of I(0), I(1) and fractionally
integrated series.
 The unit root test results as summarily represented in table 3 above show that the
series have different orders of integration [ both I(0) and I(1)].
 The above notwithstanding, the underlying procedure for the Johansen
cointegration test as well as the ARDL bounds test would be considered. We will
discuss later how to deal with this situation [I(0) and I(1)] in a multivariate
framework since the Johansen test is no longer valid. Recall that, in a single-
equation case, we use the Bounds cointegration test.

4.2.1 Johansen Cointegration Test

 Johansen co-integration test uses two test criteria namely; the trace statistic and
the maximum eigenvalue, both generated with ML technique.
 The null hypothesis is that there is no cointegration between/among the series
under examination.
 If co-integration test is carried out and the result shown with trace statistic and
maximum eigenvalue suggests that the null hypothesis of no co-integration
should be rejected, then, it implies that co-integration exists and as such, there is
long run relationship between/among the variables in the model.
 If otherwise, there is no long run relationship in the model.
 Thus, a model for cointegrating equation may be specified in line with Johansen
(1988) and Johansen and Juselius (1990) as below:

k
Z t  C   i Z t i  Z t 1  t (1)
i 1

where
p p
   Ai  I and i    A j (2)
i 1 j  i 1

9
 If the coefficient matrix  has reduced rank r  n, then there exist n  r matrices

of  and  each with rank r such that     and   yt is stationary.

 r is the number of cointegrating relationships, the elements of  are known as

the adjustment parameters in the vector error correction model and each column
of  is a cointegrating vector.

 It can be shown that for a given r , the maximum likelihood estimator of 

defines the combination of yt 1 that yields the largest r canonical correlations of

yt with yt 1 after correcting for lagged differences and deterministic variables

when present.
 Johansen and Juselius (1990) specify two likelihood ratio test statistics to test for
the number of cointegrating vectors. The first likelihood ratio statistic tests for
the null hypothesis of exactly r cointegrating vectors against the alternative r + 1
vectors is the maximum eigenvalue statistic.
 The second statistic tests for at most r cointegrating vectors against the
alternative is the trace statistic. Critical values for both test statistics are tabulated
in Johansen-Juselius (1990).
 However, where there are conflicting results between the two tests [which is not
impossible when dealing with small samples], the trace test should be used (see
Lüutkepohl et al, 2001).

Table 4: Cointegration Test Results

Trend assumption: Linear deterministic trend (Constant without Trend)
4A: Unrestricted Cointegration Rank Test (Trace)
Hypothesized Trace 0.05
No. of CE(s) Eigenvalue Statistic Critical Value Prob.**
None 0.053 14.992 15.495 0.059
At most 1 * 0.026 4.900 3.841 0.026
Trace test indicates no cointegration at the 0.05 level
* denotes rejection of the hypothesis at the 0.05 level
**MacKinnon-Haug-Michelis (1999) p-values
4B: Unrestricted Cointegration Rank Test (Maximum Eigen value)

10
 Hypothesized Trace 0.05
No. of CE(s) Eigenvalue Statistic Critical Value Prob.**
None 0.053 10.092 14.265 0.206
At most 1 * 0.026 4.900 3.841 3.841
Max-eigenvalue test indicates no cointegration at the 0.05 level

* denotes rejection of the hypothesis at the 0.05 level

**MacKinnon-Haug-Michelis (1999) p-values

Table 5: Cointegration Test Results

Trend assumption: Linear deterministic trend (Constant with Trend)
5A: Unrestricted Cointegration Rank Test (Trace)
Hypothesized Trace 0.05
No. of CE(s) Eigenvalue Statistic Critical Value Prob.**
None 0.055345 15.40656 25.87211 0.5409
At most 1 0.026440 4.930469 12.51798 0.6064
Trace test indicates no cointegration at the 0.05 level
* denotes rejection of the hypothesis at the 0.05 level
**MacKinnon-Haug-Michelis (1999) p-values
5B: Unrestricted Cointegration Rank Test (Maximum Eigen value)
Hypothesized Trace 0.05
No. of CE(s) Eigenvalue Statistic Critical Value Prob.**
None 0.055 10.476 19.387 0.568
At most 1 0.026 4.930 12.518 0.606
Max-eigenvalue test indicates no cointegration at the 0.05 level

* denotes rejection of the hypothesis at the 0.05 level

**MacKinnon-Haug-Michelis (1999) p-values

4.2.2 ARDL Bounds Cointegration Test

 The Bounds co-integration approach can also be used to test for the existence of
long run equilibrium regardless of whether the underlying variables are I(0), I(1)
or fractionally integrated.
 It involves just a single-equation set-up, making it simple to implement and
interpret.
 It is however important to point out that the test is not applicable to multivariate
models such as the VAR and its extensions.

11
  Nonetheless, we still demonstrate how to deal with the test here.
 According to Pesaran and Pesaran (1997) and Pesaran and Shin (1998), the
underlying test equation for the Bounds test is the Autoregressive Distributed

Lag  p, q1, q2, ..., qk  and can be written as follows:

k
  L , p  y t   0    i  L , q i  xi ,t  t (4)
i 1

 Where  0 is a constant, yt denotes the dependent variable, L is a lag operator,

xi ,t is the vector of regressors (where i = 1, 2,…, k) and  t is the disturbance term.

 In the long-run, we have yt  yt 1  ...  yt  q and xi ,t  xi ,t 1  ...  xi ,t  q .

th th
 xi ,t  q denotes q lag of the i variable.

 The long run equation can be written as follows:

k
y t      i x i ,t   t (5)
i 1

 Asymptotic critical values are obtained from Table CI(ii) Case II: Restricted
intercept and no trend for k = 3 (Pesaran et al. 2001).

 Decision Criteria:
 Pesaran et al. (2001) provides bounds (lower and upper bounds) on the critical
values for the asymptotic distribution of the test statistic.
 If the computed test-statistic falls below the lower bound, it implies that there is
no cointegration.
 If the test statistic exceeds the upper bound, then, there is cointegration.
 However, if the test statistic falls between the bounds, the test is inconclusive.

Table 6: Bounds Test Results

Critical Value
Dependent Wald-statistic Conclusion Lower Upper
Variable Bound Bound
Log(ASI) 2217.6 [0.000] Cointegration 1% 3.65 4.66
Log(TBR) 291.4 [0.000] Cointegration 5% 3.15 4.08
10% 2.79 3.67
Note that in each case, the second variable is considered as a regressor.

12
Based on the computed statistics and critical values provided, we can conclude that
there is cointegration in both cases.

 These are the relevant preliminary analyses for Time Series Multivariate Models.
 Let us now consider the next stage of the estimation procedure which is the main
Estimation.

13
 II. Estimation Proper:

This can be partitioned into two:

 Model Specification
 Estimation and Results

1.0 Preamble
Before we proceed to the estimation proper for the Multivariate Models, let us digress a
bit for the purpose of clarity and to clear some misconceptions about single-equation
models:
 Let us assume a single-equation model (i.e One dependent variable and one or
more independent variables);
 If all the series are I(0), and hence stationary, we can estimate the model in their
levels, using OLS estimation, for example;
 If all the series are integrated of the same order (e.g., I(1)), but they
are not cointegrated, we can just difference each series, and estimate a standard
regression model using OLS. In this case, we are only able to estimate the short
run since there is no long run;
 If all the series are integrated of the same order, and they are cointegrated, two
types of models can be estimated: (i) Long run model where the series are
estimated with OLS in their levels; and (ii) Short run model which is referred to as
“Error-Correction Model (ECM) and is estimated with OLS.
 However, if the order of integration is mixed [combination of I(0) & I(1)],
The ARDL / Bounds Testing methodology of Pesaran and Shin (1999) and
Pesaran et al. (2001) should be considered. Some of the attractions of the
methodology are: (i) It can be used with a mixture of I(0) and I(1) data; (ii) It
involves just a single-equation set-up, making it simple to implement and
interpret; and (iii) The dependent and the independent variables can be assigned
different lag-lengths in the model.
 Note that all these models also require the pre-tests previously discussed; that is,
unit root and cointegration tests.

14
 2.0 Specification of Multivariate Models

Let us consider a bi-variate model involving stock market and money market
based on our preliminary analyses. The specifications are rendered as follows:

 Vector Autoregressive Models (VAR):

k k
ASI t   0   1i ASI t i    2iTBRt i  1t
i 1 i 1
k k
(5)
TBRt   0   1i ASI t i    2iTBRt i   2t
i 1 i 1
This model is considered if the two series are stationary; that is, they are I(0) series.

 Vector Error Correction Models (VECM):

k k
ASIt  0  1i ASIt i  2i TBRt i  1ECMt 1  1t
i1 i 1
k k
(6)
TBRt  0  1i ASIt i  2i TBRt i  2 ECMt 1  2t
i 1 i 1
This model is considered if the two series are non-stationary (that is, they are I(1)
series) but are cointegrated based on the Johansen Cointegration test.

 VAR-in-First Differences:

k k
ASI t   0   1i ASI t i    2 i TBRt i  1t
i 1 i 1
k k
(7)
TBRt   0   1i ASI t i    2i TBRt i   2 t
i 1 i 1
This model is considered if the two series are non-stationary (that is, they are I(1)
series) but are not cointegrated based on the Johansen Cointegration test.

15
  Toda-Yamamoto Model:

k k  dmax k k  dmax
ASIt  0  1i ASIt i   2j ASIt  j  1iTBRt i   2j TBRt  j  1t
i 1 j k 1 i 1 j k 1
k k  dmax k k dmax (8)
TBRt  0   1i ASIt i   2 j ASIt  j  1iTBRt i   2 jTBRt  j   2t
i 1 j  k 1 i 1 j k 1

This model is considered if the two series are of different orders of integration (say
I(0) and I(1) series).

Definition of Relevant Terms

(i) k denotes the optimal lag length. This is determined by the usual
information criteria such as AIC and SIC.
(ii) dmax is the maximum order of integration. For example, if ASI is I(0) and TBR
is I(2); then, the maximum order of is 2. That is, d max  2 .

16
3.0 Estimation and Results

Let us now demonstrate how these models can be estimated using Eviews.

 Estimation of VAR
Step 1: Determine the Optimal Lag Length  k 

17
18
You will observe that the various information criteria suggest that we should have a
maximum lag length of 4 for each variable.

Step 2: We can now estimate a bi-variate VAR(4) Model:

19
 Estimation of VECM:

20
 Estimation of VAR-in-First Differences

21
 Estimation of Toda-Yamamoto:

22
Note that it is practically impossible to interpret the results of multivariate models.
We estimate these models in order to be able to test for Causality and compute
Variance Decompositions and Impulse Response Functions.

Before we proceed to Causality Test, Variance Decomposition & Impulse Response

Functions, we have to verify that the estimates of the chosen multivariate model are
reliable. This will require diagnostic checks/Post-estimation tests.

The most relevant post-estimation test for Multivariate Models is the Serial Correlation
test (using the LM test).

Note that based on our preliminary Analyses, the appropriate multivariate model is the
Toda-Yamamoto (TY) model. Therefore, our diagnostic tests would make use of the
estimates of the TY model.

From the TY estimation output, follow the steps below for the LM test:

23
Note that the null hypothesis for the test is that there is no serial correlation. The result
presented above suggests that there is no serial correlation. If there is evidence of serial
correlation, it is better resolved by increasing the optimal lag length.

3.1 Granger Causality Test:

In a simple definition, using our series, ASI is said to Granger-cause TBR if TBR can be
better predicted using the histories of both ASI and TBR than it can by using the history
of TBR alone.

There are three possibilities:

 Unidirectional Causality where only one of the groups is statistically significant.
 Bidirectional causality where at least two groups are statistically significant.
 Non-causality where none of the groups is statistically significant.

Note the following when dealing with Granger causality:

 The chosen multivariate model must be well specified. That is, there
should be no evidence of significant serial correlations in the residuals.

24
  Rejection of the null hypothesis implies rejection of Granger non-causality.
In other words, the alternative hypothesis is that there is Granger
causality.
 The Wald test that is chi-square distributed is used to test for Granger
non-causality.
 If there is cointegration; then, there must be Granger causality either
unidirectional or bidirectional. This is a good cross-check on the validity
of the causality results.
 In other words, if there is cointegration and none of the groups is
statistically significant implying no causality; then, there is a problem. If
there is long run relationship, it must also reflect in the causality results.
One may have to check the sample size of the data. It may be too small to
satisfy the asymptotics underlying the cointegration and causality tests.
 When testing for Granger non-causality, the model must be expressed in
levels and not in first differences. Note that only VAR and TY models are
expressed in levels and are therefore used when testing for Granger non-
causality.
 In other words, the VEC and VAR-in-First Differences must not be used
when testing for Granger non-causality.
 Nonetheless, these models can be used for purposes other than testing for
Granger non-causality such as Variance Decomposition and Impulse
Response functions.
 Teaser: Since we cannot perform Causality test in First differences, how
do we then deal with series that are non-stationary but have the same
order of integration [say I(1)]?

Testing for Granger non-causality:

 Using the VAR Model:

 TBR is said to Granger-cause ASI if  2i  0 ; otherwise it does not.

 ASI is said to Granger-cause TBR if 1i  0 ; otherwise it does not.

 Using the TY Model:

 TBR is said to Granger-cause ASI if 1i  0 ; otherwise it does not.

 ASI is said to Granger-cause TBR if  1i  0  ; otherwise it does not.

25
Again, based on the preliminary tests, the appropriate model for the Granger non-
causality test is the TY model. Nonetheless, we shall consider the two relevant models
for the test.

 Testing for Granger non-causality using the VAR model:

Step 1: Estimate the appropriate VAR Model

Step 2: Use the VAR estimates to test for causality:

26
  The result indicates that there is no causal relationship between the two series.

 Testing for Granger non-causality using the TY model:

Step 1: Estimate the appropriate TY Model:

 The result indicates that there is no causal relationship between the two series.
 Recall that the ARLD Bound testing cointegration test indicates the presence of
long run. However, the Wald test is saying otherwise. You may consider the
following to resolve the conflict:
5 Increase the sample size either by expanding the data scope or by considering a
higher data frequency.
6 The results may be sensitive to structural breaks. Therefore, there may be need to
test for common breaks in the VAR model.
7 It may also be necessary to consider relevant intervening variable(s) justifying the
expected long run relationship. That is, explore the appropriate transmission
mechanism through which one series can Granger-cause another.
8 Check the measurement of variables. For example, if investors consider returns
when making investment decisions; then, returns of the series must be used for
modelling and not the index form.
9 Also, it may be more appropriate to model in real terms if inflation is not captured
as one of the intervening variables in the multivariate model.

27
3.2 Impulse Responses and Variance Decompositions

3.2.1 Impulse Responses

Impulse response functions (IRFs) show the effects of shocks on the adjustment path of
the variables in the VAR model. IRFs can also be graphically presented showing the
effect of shocks on the current and future path of the variables under consideration. In
essence, IRs show how these variables react to different shocks in the model.

Recall the generalized VAR model:

= + + +⋯+ +
This model can be written equivalently as:
= + (9)
where = − − − ⋯−

We can write equation (9) in a more compact form as:

∞

= + (10)

= + + + +⋯ (11)

Equation (9) which is usually used to trace the impulse response is regarded as the
Vector Moving Average (VMA) model. The impulse response functions, that is, the
effects of the various shocks of the explanatory variables on the dependent variables can
be determined by differentiating equation (9) with respect to each of the shocks
( , … , ).

3.2.2 Forecast Error Variance Decomposition

Forecast error variance decompositions (FEVDs) measure the contribution of each type
of shock to the forecast error variance. FEVD tells us how much of a change in a
variable is due to its own shock and how much due to shocks to other variables. In the
SR most of the variation is due to own shock. But as the lagged variables’ effect starts
manifesting, the percentage of the effect of other shocks increases over time.

For example, from equation (11), we can calculate the n-period forecast error of :

Start from 1 period:

= + + + +⋯

28
 = + + +⋯

1-period forecast error: − =

Proceed in the same way and get 2-period forecast error:

− = +

3-period forecast error:

− = + +

Therefore, n-period forecast error:

− = + + +⋯+ =

Considering for example, the variance of its n-step-ahead forecast error can be
expressed as:

Total Var ( )= Proportion of Variance Due to shock

+ Proportion of Variance Due to shocks of other variables in the

model

Note the following about the FEVDs:

 Proportion of Variance Due to own shock decreases over time

 Proportion of Variance Due to shocks of other variables in the model increases
over time
 If errors of other variables can explain none of the forecast error variance of the
series under consideration say at all forecast horizons, then is assumed to
be exogenous.
 However, if errors of other variables can explain most of the forecast error
variance of the series under consideration say at all forecast horizons, then
is assumed to be exogenous.

Both IRFs and FEVDs computations are useful in assessing how shocks to economic
variables reverberate through a system.

29
Computation of Impulse Responses and Variance Decompositions using EViews

 Click Quick and choose Estimate VAR option

 Enter the names of endogenous and exogenous series in the appropriate edit
boxes.
 Here, we have listed log(ASI) and log(TBR) as endogenous series and C as the
constant exogenous term.

30
  Click OK on the VAR specification box and VAR estimates would appear as
follows:

 Impulse Response Function (IRF)

 From the VAR(4) model results, CLICK View/Impulse Response

31
 The impulse responses as shown below, indicate that stock market, which is
captured by all share index (ASI), responds negatively to money market shock.
 Also, money market which is proxied by treasure bill rate (TBR) responds
negatively to stock market shock.
 The markets however, respond positively to own shocks resulting from
innovation within the respective market.

 Forecast Error Variance Decomposition

 From the VAR Results/View/Variance Decomposition

32
 The variance decomposition results allow us to make inference over the
proportion of movements that is due to a market own shocks versus shocks to
other variables in the system.
 For instance, the variance decomposition results reported within a 10-year
horizon as presented below shows that 99.3 per cent of innovations in stock
market are explained by its own past values, yet only 0.69 per cent of the
innovations is due to shocks to money market.
 Also, only 2.5 per cent of innovations in money market can be attributed to
shocks due to stock market.

33
 VECM/Impulse Response Function/Variance Decomposition
 Similar to the procedure for VAR, CLICK/Quick/Estimate VAR.
 From the VAR Type box, choose the second option (i.e. Vector Error Correction).
 From the basic option, CLICK/Cointegration and indicate the number of
cointegrating equations (for the purpose of class demonstration; choose 1).

 Similar to the VAR procedures, the IRF and FEVD would be performed on the
VECM estimates.

34
 From the VECM result CLICK/View/IRF
 Also from the VECM result CLICK/View/Variance Decomposition

35
36
 Toda-Yamamoto/Impulse Response Function/Variance Decomposition

You specify the model based on the order of integration of the series:

Response to Cholesky One S.D. Innovations ± 2 S.E.

Res pons e of D(LOG(ASI)) to D(LOG(ASI)) Res pons e of D(LOG(ASI)) to LOG(TBR)
.10 .10

.08 .08

.06 .06

.04 .04

.02 .02

.00 .00

-.02 -.02
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10

Res pons e of LOG(TBR) to D(LOG(ASI)) Res pons e of LOG(TBR) to LOG(TBR)

.3 .3

.2 .2

.1 .1

.0 .0

-.1 -.1
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10

37

View publication stats