|
PureBytes Links
Trading Reference Links
|
There is a new econometric method called Cointegration which measures the
relationship between multiple data series. Contegration is still a bit new
and "exotic" but it allows you to pump in X data series and it outputs how
many of these data series are cointegrated or have a statistical
relationship. The best Cointegration technique developed so far is by a guy
called Johansen (about 10 years ago) whose technique is called Johansen
Cointegration. The output is not a correlation but instead a level of
confidence or probability. Like all things that are econometric in nature
(e.g. linear and multiple OLS regressions) they are good for modelling
historical data but are virtually useless on real-time analysis. I know a
fund with a well renowned academic that have tried to use Cointegration in
real time trading but since Feb 2002 they have not posted their results. I
do not know what has happened to them. Things started getting juicy in
equities from March/April onwards. Please don't ask me their names.
Using Ordinary Least Squares (i.e.OLS) multiple regression analysis assumes
the data you are using has constant variance (which most financial data
does not, that is why ARCH and GARCH was invented and they are useless
anyway) and the other assumption is that your residuals from the regression
are normally distributed which generally they are not. If you actually run
the output of these models you will find that most OLS regressions produce
invalid output that are outside the assumptions of the model. In other
words, watch out if you use them. Most things in econometrics have moved
beyond multiple regressions though. OLS linear/multiple regression is
taught to undergraduates and post-graduates because it is the simplest
things to teach. Just because it is taught it does not neccessarily mean
that it should be used especially for real time trading.
HTH
Robert Bianchi :-)
r.bianchi@xxxxxxxxx
===============
|