Financial Econometrics ECO373 – Assignment

Abhay Italiya – 2110110021

B.Sc. Research in Economics and Finance

Q1. Consider the data file named SIMS_2.xls. Use the first series ‘Y3’ to perform the
following tasks:

a) Plot the sequence against time

The following is the graph depicting time series for Y3_

b) Plot the ACF and PACF of the above series. Compare the obtained
sample ACF and PACF to those of a theoretical AR (2) process.
c) Estimate the series as an AR (1) process. How you would check for
model adequacy?
We can check for the model adequacy through analyzing the correlogram of the
residuals of the estimated AR (1) Model_

We can apply residual analysis and SIB-AIC analysis to check whether the AR(1) fits
the model or is not a good fit for the same.

1. Residual Analysis – We can see from the residuals of ACF and PACF from the
below screenshot that there are significant autocorrelations beyond lag 1 following
through lag 2 , lag 3 and so on. These significant autocorrelations indicate
inadequacies in the model.

2. AIC-SBC test – We can see that the Alkalike Info Criterion and Schwarz Criterion
are 0.716469 and 0.794624 respectively. They are significantly higher than they
should be and are not near to 0. The SBC and AIC being between 0.4 and 0.3 is a
good fit for the model which is not shown in the above case.

Hence, we can conclude that the above AR(1) process is not so good fit for the
model and is less adequate than the other.
d) Could an ARMA (1, 1) process generate the type of sample ACF and PACF
found in part (a)? Estimate the series as an ARMA (1, 1) process. Is this model
adequate? Please explain.

In an ARMA(1,1) process, the ideal Autocorrelation Function (ACF) shows a positive

correlation at lag 1 due to the autoregressive component, followed by a rapid decay to
near zero for subsequent lags. The Partial Autocorrelation Function (PACF) displays
a non-zero value at lag 1, representing the autoregressive influence, and a quick drop
to near zero for lags beyond 1.
Again, just as from part c we could use the correlogram of the residuals of the
estimated model in this case particularly the ARMA (1.1) model _

If you have the residual correlogram of an ARMA (1,1) model, you may use it to
evaluate how well the model fits the data. Here are some important things to think
about:

1. You can evaluate that the ACF of residuals to show no significant

autocorrelation at any lag beyond lag 1.
2. You can also see that the PACF of residuals will not show any significant or
major spikes at any lag beyond lag 1.

We see these criteria are seen resembling in our ARMA (1,1) model hence it fits the
model adequately. This model passes the residual analysis.
Now, after analyzing the SBC and AIC to be 0.686959 and 0.582752 we can say that
they are fitting the model adequately.

e) Which model, according to you, fits the underlying data best? Please
explain your answer.

Considering our syllabus there can be multiple possibilities of processes that can fit
the model which are – AR(1) , AR(2) , AR(3) , AR(4) , MA(1) , MA(2) , MA(3) , MA(4),
ARMA(1,1) , ARMA(1,2) , ARMA(2,1) , and ARMA(2,2)

After evaluating an analysing each of the possibilities for the processes we came to
know that AR(2) process fits the model best in my opinion with a very low
autocorrelation for lags beyond lag 1 for ACF and PACF preventing inadequacies
through the correlogram of residuals of the model. Furthermore, the very low SBC
and AIC closer to 0 indicate the same.

Screenshots and Explanation for the same_

As explained above we can see through the residual analysis that_
The autocorrelation between the residuals for the ACF and PACF is decreasing after
lag 1 and there are no significant autocorrelation leaving no space for inadequacies.

We can also see through the SBC and AIC being 0.580232 and 0.476026 which are
much smaller and closer to 0 thus indicating the AR (2) is the best process that can
fit the model. Lastly, it also has the lowest standard error of 0.299907.

Additionally, since the constant's T-statistics value is greater than 4, we may observe
that AR (1) and AR (2) have statistically significant coefficients. Furthermore, the
coefficient's standard error is negligible.

Increased parameter testing demonstrates that AR(2) fits the data better; R-squared
and adjusted R-squared are both less than those of any AR/ARMA process.
Q2. Consider the data file named QUARTERLY.xls. Use the series PPI to
perform the following tasks.

a) Plot the series PPI and growth rate of PPI

b) Comment on the time plots.

Looking at the time plot of ppi series seems to have an upward trend which
increases over the period of time as the majority portion of ppi has a positive line
graph showing that the data is having an upward trend. The drift can represent an
expected average return or growth rate, while the random walk component captures
the short-term volatility or unpredictability.
On the other hand, a pattern of variations around a mean is shown in the growth rate
of PPI, which is calculated by calculating the first difference of the logarithm of PPI.
Given that the growth rate shows no systematic pattern over time, this feature is
suggestive of a stationary process. Notably, a popular technique to create
stationarity in a time series is the first differencing of the log of PPI.

c) Fit an appropriate ARMA model to the growth rate of GDP series. To select
among alternate models, use AIC, SBC and Ljung-Box Q-statistic as the model
selection criteria.
Estimation of ARMA (1,1) Process and Correlogram of Residuals.
We can clearly observe that SBC and AIC are -6.614637 and -6.689018 respectively.
Both SBC and AIC are very lower than 0 which indicates that the process is a good
fit of the model for the same.

A considerable autocorrelation in the residuals, as indicated by the Ljung-Box test,

may indicate that the selected model is insufficient. It may be necessary to look at a
more intricate model or look into the autocorrelation's origin.

Q3. Consider the data file named MONEY_DEM.xls. The series RGDP series
denotes real GDP of the U.S. economy from 1959 Q1 to 2001 Q1.

a) Plot the sequence against time. Does the series appear to be non-
stationary? What kind of nonstationary would you expect? Give an intuitive
explanation.
From the part c of Q3 and intuitively thinking rather than the random walk plus drift
with the Dickey-Fuller test is indicative of the absence of a unit root, implying that the
time series is stationary or trend-stationary, depending on the model specifications.

b) Plot the first difference of log real GDP (call it: fdlrgdp). Fit an ‘appropriate’
ARMA model to this new series (fdlrgdp).
Estimation of ARMA(1.1) process which best fits the model and correlogram of
residuals-
We can clearly observe from above that_
1. Firstly, through the residual analysis that the autocorrelation between the ACF and
PACF residuals is very low or near zero beyond lag 1 through the correlogram thus
showing that the ARMA (1,1) is a good fit process for the model.
2. The ARMA (1,1) has the lowest of SBC which is -6.567414 and an AIC of -
6.641794 thus indicating it’s the best fit for the model among other processes which
were ARMA (2,1), ARMA (1,2) and ARMA (2,2).

c) Using the ADF test, test for the presence of unit root process in the series
log real GDP.

Unrestricted Model: tau-tau

Here the t-statistic value is -2.854557 which is inside the acceptance region thus, we
fail to reject the hypothesis. (value is lower than the critical values at particular given
significance levels)
Restricted Model: tau-mu

Here too like before the t-statistic value -0.651090 which is inside the acceptance
region and thus we fail to reject the hypothesis. (value is lower than the critical
values at particular given significance levels)

Restricted Model- tau

Here the t-statistic value is 6.574527 which is outside the more than the critical
values at the particular significance levels and thus we reject the hypothesis.

d) Does your test result confirm or contradict your expectation from part (a)?

My test results confirm my expectation from the part (a) , as after doing the augmented
dickey fuller test we know that there is a unit root and hence we fail to reject the null
hypothesis and thus the time series is non-stationary.