Traders' Roundtable

I've heard it said that "cointegration" analysis has replaced correlation analysis in many financial communities.

Does somebody have a clear and concise argument for the switch?

How about a software package that does this??

Thanks in advance

FT

I don't, but the Wilmott forums do.

try Wilmott in google and sign up for the forums, its a quantative finance website.

In 'The Quantitative finance project FAQs' thread there is a short description of cointegration

[/url]

Do a google on cointegration and financial.

A definition can be found at http://economics.about.com/library/glos ... ration.htm and I would reproduce it if I understood it.

Paul Wilmott didn't think "cointegration" was important enough to include in his book http://www.amazon.com/exec/obidos/tg/de ... 471498629/

Cointegration seems to be a binary indicator: a vector of time series are cointegrated, or they aren't. It doesn't sound all that useful to me but then I'm a practitioner not an academic.

One can google around and find a number of interesting papers, most of them econometric, and many of them directly related to markets (e.g. currencies, metals, interest rates, ags). The CATS and RATS software pops up in a number of papers and places, as do some other implementations and algorithms. The most accessible application seems to be building non-cointegrated portfolios (in comparison to non-correlated portfolios). I also found a few cautionary findings that some of the cointegration algorithms can fail to either establish cointegration when it exists, or mis-identify cases where it doesn't. I don't know enough about the field to discern the issues.

What I did find interesting is that a number of other trading ideas seem to pursue similar paths of differencing a price series and factoring for noise. (the focus of cointegration is on the consistency of residuals, or shape of noise as a result of the differences). For instance, a 2004 TASC article looked at trend identification using Q and B vectors. In particular, it seemed to do an excellent job to me of identifying the end of a trend (I thought it lagged to much at detecting onset of trend). So, that might be another application.

Cheers,

Kevin