Marta

Montesano 1517275
Ilaria Nava 1496074
Stefano Tripodi 1529098

PROBLEM SET 3

Exercise 1

In order to estimate a VAR on the three series (real_GDP, price_deflator,
fed_fund), we decide to keep the three processes undifferenced, despite the fact
that two series present a unit root and the third one ,price_deflator,displays a
deterministic time trend .
Supporting the recommendation of Sims, Stock and Watson (1990) against
differencing, even if variables contain a unit root, we argue that differencing
throws away information concerning the comovements in the data and reduce
the quality of the analysis of the interrelations among the variables.
We start by estimating a VAR (1) model. Since the estimated residuals fail to pass
the Q-test for white-noiseness, we proceed considering other VAR specifications
of superior orders.
Finally, we conclude that the model which displays white noise errors in all its
equations is a VAR (4).
To check the quality of our model, we compare also the results of the BIC
criterion (generalized in the multiequation case ) for the specifications
VAR(1)...VAR(4). The statistics suggests us that a good balance between
parsimony and goodness of fit would be reached with a VAR(2) model.
Realizing the likelihood-ratio test for the comparisons of the same specifications
considered before, we would say that the specification that fits best our data is,
instead, a VAR(3).
Despite these results, we prefer to use a model that contains a sufficient number
of lags p to ensure that the residuals in each of the equations are white noise, and
that the model doesn t face problems of mispecification. Thus, our final choice is
a VAR(4).

Once we have estimated the standard form of our VAR (4) model, we control that
the process is actually stationary computing the eigenvalues of the four
coefficient matrices of the model: they all result to be smaller than one in
absolute value, so stationarity is confirmed.
Finally, using the Choleski decomposition on Q (var-cov matrix of the reduced
form model residuals) and constraining the to be an identity matrix, we are
able to identify the structural model and the monetary shock effects by the
analysis of the IRF.

So we generate N random integers from 1 to 204, in order to sample with
replacement from the sample distribution of the VAR(4) residuals ; using the
estimated coefficient from the reduced model (contained in the matrixes A1,
A2, A3, A4) and premultiplying the new residuals series for the Choleski
factor, we construct a newly simulated sample of the three variables
real_GDP, price_deflator, fed_fund:
y1(i,:)=y1(i-1,:)*A1'+y1(i-2,:)*A2'+y1(i-3,:)*A3'+mi_1(i,:).
Then we estimate a VAR (4) model on this new series following the process
explained above and then we write it in its Companion form.
Afterwards we set a unitary monetary shock (shock=[0 0 1]') and compute the

Marta Montesano 1517275

impulse response function for the three new series, exploiting the VMA
representation of the VAR(4) companion form itself.

At this step of the process, we run also a Forecast Error Variance Decomposition
in the matlab file.

Finally we repeat all the steps just presented for computations for the impulse
response functions for 1000 times using the Bootstrap procedure, obtaining
1000 IRF.
Plotting the 2.5% and 97.5% percentile of that empirical distribution, we get the
95% confidence bands for the impulse response functions.

Impulse response function confidence interval for price deflator to a monetary shock
0
-0.05
-0.1

10

12

14

16

18

20

Impulse response function confidence interval for GDP to a monetary shock

10

12

14

16

18

20

Impulse response function confidence interval for fed fund rate to a monetary shock
2
1
0

10

12

14

16

18

20

As we can see from the graph, in response to a unitary monetary shock we
observe an initially slow dimishing path for prices (deflation), followed by a
period of slow increase of prices ; we obtain also an initial contraction of GDP,
followed by a recovery in the long run. We underline that the coefficients are to
be interpreted as the percentage deviation from the year before. These results
are consistent with Macroeconomic theory.

Exercise 2

Marta Montesano 1517275

In this exercise we are required to identify and estimate the components of
productivity and hours variations that are associated with technology and non
techonology shocks.
For this reason we proceed applying a similar procedure to the one presented in
exercise 1: we start by estimating the right model for our data (considering the
evaluations we will made on the stationarity of the series) and then we will use
the triangular decompositon method in order to identify the structural model
parameters and to compute the IRF related to monetary and technology shocks.
Lets proceed so with order.
First of all we control that our series (x,n,gdp=x+n)are stationary by the mean of
an augmented Dickey Fuller test. Since all the three variables present a unit root,
we take their first difference values, and check again their stationarity with the
ADF test.
We will work, therefore, using the series in difference X=diff(x) N=diff(n) as
our variables of interest.
As first attempt, we estimate a simple VAR(1) model on the series Y=[X N]. We
obtain non white noise residuals, thus we go on with other specifications.
The second attempt we make is to estimate a VAR (2) model on Y; this time we
obtain normal and non autocorrelated residuals, and so we conclude that this is
the best specification for our data. Moreover we obtain evidence of the fact that
the eigenvalues of the matrices of the VAR (2) model lie within the unit circle and
so that the latter VAR model is stationary as well.
As second step, we proceed computing the responses of the system to a unitary
non -technology shock.
Once decomposed la Cholesky the var-covariances matrix of the residuals (Q)
of the reduced form of the model, we build the companion form of the VAR(2)
model appropriately, and estimate the response function of the productivity in
difference ,of the hours of labor in difference and of the GDP.
We proceed by doing the same thing for the technology shock and finally we run
a bootstrap in order to obtain the confidence bands for the impulse response
functions. The results are shown in the graph below.

Marta Montesano 1517275

First we consider a non technology shock. As we can see from the graph, there is
a sharp decrease in the productivity series, followed by a recovery. Then the
series stabilizes at a level higher than the initial one. For output and hours series
we observe a decline.
Next, looking at the technology shock, we notice that both productivity and
output rise and then decrease, stabilizing at a lower level. For hours we observe
only a slight increase, but the effect is neglegible.