Scribd
Upload
5 views34 pages

Cointegration and VECM

Uploaded by

nicebanach3
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
5 views34 pages

Cointegration and VECM

Uploaded by

nicebanach3
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 34

Cointegration and VECM

Cointegration
• Generally, if both the series in a bivariate VAR are stationary i.e. both
are I(0), then the VAR is stationary. And if a VAR is stationary, it can
be used for forecasting.
• If a VAR is non-stationary, then we check for the stationarity of
individual series.
• For a bivariate model, if one series is stationary i.e. I(0) and the other
is not i.e. it is I(1), we take the first difference of the I(1) series to
make it stationary. This would make it an I(0) series.
• Once both series are stationary (one is already I(0) and the other has
been differenced to become I(0)), we can apply a VAR model.
Cointegration
• VAR model in this setup can indeed be used for forecasting, even
when one series is in differenced form.
• When we use the VAR on the stationary (differenced) series, it will
produce forecasts for the differenced values of the I(1) series.
• For the I(0) series (the one that is already stationary), the VAR will
directly produce forecasts of its actual values.
• Since the forecast for the I(1) series is in differenced form, we need to
reintegrate it to return to the original level. To do this, we start with
the last observed level of the I(1) series before the forecast period
and add the forecasted differenced values cumulatively. For instance:
Cointegration
Cointegration
• If both the series in the VAR are I(1), then we can transform the variables,
taking first differences, so that they are stationary.
• When we take first differences, we lose a lot of important information. We
might know how a variable is changing; but we don’t know its actual value
(its level).
• We are not modelling the variables we are really interested in, but their
rates of change.
• Similarly, when we model the difference between two variables, we are
estimating the statistical properties of a new variable, not the original two
variables we are interested in.
• While this is better than nothing, Engle and Granger (1987)showed how we
can do much better than that.
Cointegration
• Econometrician Michael Murray (1994) wrote a humorous, but very
useful piece, linking cointegration with a drunk person walking his
dog.
• A drunk staggering out of a bar follows a random walk. An unleashed
dog might also follow a random meander.
• But if the drunk staggers out of the bar with his dog, even without a
leash, the distance between the drunk and his dog will be relatively
stable.
Cointegration
• If the dog meanders too far away, the drunk will randomly call out to her
dog, and it will move closer to its owner.
• Sometimes the dog will randomly bark for its owner, drawing her closer to
it.
• If you see the drunk, the dog will not be too far away. Likewise, if you see
the dog, its drunk owner should be nearby. The barking and calling—and
the staggering toward each other—is the error correction mechanism,
whereby when the two diverge, they begin to converge again.
• We might not know where they’ll go as a pair, but we can be fairly certain
they’ll be close to each other. That is the essence of cointegration.
Cointegration
Consider a new set of variables X and Z
• They are both integrated i.e. both are I(1) – non stationary.
• But are they cointegrated?
• The gap between them is increasing; that is, their difference is not
stationary, so at first glance they do not seem to be cointegrated.
• But there does seem to be a relationship between X and Z. Ignoring
their different slopes, when X dips so does Z. When X spikes so does
Z. The problem is with their slopes.
• What would make the difference between X and Z stationary is if we
could either tilt X up or tilt Z down?
Estimating the ECM
(If there are more than 2 series, we call it VECM)
VECM for the US Life Expectancy Dataset
FORECASTING USING VECM

You might also like