ECON3206

Financial Econometrics

Tutorial 4 (Week 6)

Tutor: Felipe Pelaio

Email: [email protected]
 Outline
 Tutorial 4.
◦ We will go through Q3, Q4, and Q6.

 Assignment 1 - Results
 Question 3 – Tutorial 4

We will do this one together on the board.

 Question 4

Notice

The data isn’t in the quarter format, but in months. So, to select the time frame
asked in the question we should use “03/01/1984 03/01/2012” (notice that the
date is also in the inverted order: month/day/year).
 Visually speaking, would you say this series
is mean stationary?

In the period 84-90, approximately, the

average inflation rate seems to be larger than
in the remaining of the series. But is it
strong enough?

The spike we see in the 00’s seems to be a

one-off event.

Let’s have a look at the local averages.

They roughly seem to be about the same

level.

We need to perform an ADF test to check if the

series is stationary or not (more rigorously, to test
for the presence of a unit root).

Recall that we need to specify which version of

the ADF before running it. There were three
options:

- No intercept or trend;
- Intercept but no trend;
- Intercept and trend.

Which one is more suitable in our case?

 Question 4
𝐻𝐻0

I believe the first option is more appropriate because the intercept is near zero. But if you believe there
is an intercept, you may use the results on the right.

Regardless of your decision (and even if you see a trend in the inflation series), the result is the same if
we adopt a 5% significance level: we reject the null hypothesis of unit root, i.e. we have stronger
evidence that our series is stationary.

Recall that this is the first step for the Box-Jenkins forecasting procedure. The next one is to look at the
correlogram of the series and try to fit a few models.
Question 4

We can see a clear exponential decay in the

ACF and two significant spikes in the PACF.

This is a typical behaviour of an AR(2) model.

So, we can start with it but remember that it’s
always a good idea to have a look at some
models with similar lags: AR(2), ARMA(1,1),
ARMA(2,1), ARMA(1,2) and ARMA(2,2).

It’s important to have the table below in mind.

Next step is to select which of these models is

the best one.
 Question 4

After collecting the AIC and SIC from each of the models we attempted, the best one seems to be
the ARMA(1,1) as it minimises the AIC (best fit).

The next step is to analyse the specifics of the model we selected.

We need to analyse the regression results,
particularly (1) the statistical significance of the
regressors of interest, (2) the magnitude of the R-
squared statistic, (3) the DW and (4) the inverted
roots.

All the regressors seem to be statistically

significant at the 5% significance level and the R-
squared seems small for time series.

All the inverted unit roots are smaller than one

(the inverted roots are all larger than one, i.e., the
roots fall outside the unit circle).

The DW stat is close to 2, suggesting no residual

autocorrelation (correct model specification) but
we need to analyse the correlogram of the residual
series to gather stronger evidence.

The residuals do not show any statistically

significant spikes in any of the lags we have
displayed so it seems to be a white noise series.

In other words, we seem to have the correct

specification (ARMA(1,1) seems to be a good
model in the sense that it captures the main
features of the data we have).
Actual and fitted series are reasonably close. Recall that the R-squared was low so they won’t be
that close together.

The residual series has a mean zero and most of the realisations fall within the 95% CI (recall that
we expect around 5% of those realisations to be outside the CI). We don’t seem clusters or trends or
breaks. We can see a couple of spikes that we could regard as just one-off events. One could try to
model them, though (how?).
On the LHS, we have the point and interval forecasts for the years 2010 until 2012Q1, whereas in
the RHS we have the comparison between what effectively happened (blue line) and what our
model predicted (red line – notice the red line before the forecasting period is both the actual and the
forecast as they’re the same).

Sometimes our forecast overestimates the actual, sometime it underestimates. That’s actually a good
sign.
Question 6

Visually speaking, they seem to move together

but not perfectly.

We can see this by looking at how in the period

prior to 1994-95 when one of the series seemed
to increase the other was doing the same. And
after this period, both were a lot closer and
seemed to move in synchrony, although there are
clear period (e.g., 00’s) during which one moves
in the opposite direction to one another.
 Question 6

The output of the regression is:

There is a positive and strong association

between the two variables in this regression
(as revealed by the estimated coefficient and
its p-value).

The p-value seems quite high as we’re using

only one variable to explain another. This
reveals that the variables “walk together” side
by side.
Recall that we can express the regression above as:

𝜀𝜀𝑡𝑡 = 𝑅𝑅𝑡𝑡 − 𝛽𝛽0 − 𝛽𝛽1 𝐼𝐼𝐼𝐼𝐼𝐼𝑡𝑡

If both 𝑅𝑅𝑡𝑡 and 𝐼𝐼𝐼𝐼𝐼𝐼𝑡𝑡 are I(1) processes, it’s likely 𝜀𝜀𝑡𝑡 will also be. And this is problematic because we
wish to have an I(0) (stationary) random disturbance.

If 𝜀𝜀𝑡𝑡 turns out to be I(0) then 𝑅𝑅𝑡𝑡 and 𝐼𝐼𝐼𝐼𝐼𝐼𝑡𝑡 are said to be cointegrated. Intuitively, cointegrated series
have a long term (likely causal) relationship. One would expect short run interest rates to be causing
(i.e., directly affecting) inflation and vice-versa. Cointegrating relationships would presumably
capture that.
If, say, the two variables above were not cointegrated then it’s likely they share a common trend,
which would explain why they’re so strongly associated.

One way to test for cointegration is to check if the residuals have a unit root, thus it could be tested
with an ADF:
∆𝜀𝜀𝑡𝑡̂ = 𝜌𝜌𝜀𝜀𝑡𝑡−1
̂ + 𝑣𝑣𝑡𝑡
where 𝑣𝑣𝑡𝑡 is an iid error term.

The question asks us to not include an intercept or trend (so same specification as above). The results
are in the table below.

The ADF test returns a very small p-value

(much smaller than a 1% significance level).
So, we may strongly reject the proposition of
a unit root in the series.

In other words, 𝜀𝜀𝑡𝑡̂ seems to be I(0), implying

that 𝑅𝑅𝑡𝑡 and 𝐼𝐼𝐼𝐼𝐼𝐼𝑡𝑡 are likely cointegrated.
The Engle-Granger test is exactly the same as the ADF test we performed in (b) but with adjusted
critical values for the fact that we are using the residuals of the initial regression (not raw data like 𝑅𝑅𝑡𝑡
or 𝐼𝐼𝐼𝐼𝐼𝐼𝑡𝑡 ).

The test can be done in Eviews by using the following: open the two series as a group and then click on
View. Then on “Cointegration Test” then “Single-Equation Cointegration Test”. Choose the option
“Engle-Granger” and click ok.

The relevant row for us is the one highlighted and we would therefore reject the null hypothesis of no
cointegration considering either a τ or z statistics.

So, although the ADF didn’t have the correct critical values (and is therefore an invalid test of
cointegration), they both concluded the same.
 Question 6

If two variables are cointegrated they will theoretically converge to a point in which deviations will be
zero (the idea is that in the short run random shocks disturb the long run relationship these two
variables have, leading them to different directions. But the long run trend will slowly overpower these
short run shocks and lead to no “deviations”). That is the idea behind the error correction model.

So, for the equation:

𝜀𝜀𝑡𝑡 = 𝑅𝑅𝑡𝑡 − 𝛽𝛽0 − 𝛽𝛽1 𝐼𝐼𝐼𝐼𝐼𝐼𝑡𝑡

To lead to zero 𝜀𝜀𝑡𝑡 (smaller because it’s now positive), the next realisation of 𝑅𝑅𝑡𝑡 (𝑅𝑅𝑡𝑡+1 < 𝑅𝑅𝑡𝑡 ) should be
smaller and the next one of 𝐼𝐼𝐼𝐼𝐼𝐼𝑡𝑡 (𝐼𝐼𝐼𝐼𝐼𝐼𝑡𝑡+1 > 𝐼𝐼𝐼𝐼𝐼𝐼𝑡𝑡 ) should be larger.
Question 6
What strikes my attention when I look at all these numbers is how most independent variables are
statistically insignificant at the 5% level for both equations (also, notice how low both R-squared stats
are), especially the one on the left.

The one we should pay special attention to is the one named “E(-1)”, which represents the
cointegration relation between 𝑅𝑅𝑡𝑡 and 𝐼𝐼𝐼𝐼𝐼𝐼𝑡𝑡 . In both cases, it is not statistically significant which is
concerning as we had evidence earlier that this relationship existed.

One of the reasons that might be causing this is the excessive amount of seemingly irrelevant variables
(recall that including redundant variables in the model is likely to lead to inflated standard errors,
which could explain the large standard errors of the cointegration terms).
The answer to this question will obviously vary depending on which variable(s) one decides to
eliminate. A proposition would be the one above.

Notice that these specification are preferable over the initial ones because both models have a smaller
AIC, are more parsimonious (I suspect they could be even more) and the cointegration relation for at
least 𝑅𝑅𝑡𝑡 seems to exist.