Thursday, January 3, 2008

Cointegrate in 2008

Successful short-term trading methods may be developed around the concept of cointegration. The core data set needed to explore these methods begins with identification of security pairings whose return series are correlated in a mean reverting manner.

The way the intuition behind cointegration was first related to me, was to imagine a man walking his dog. The man and dog may seemingly move independently and in different directions for a while, but the leash ensures that they will eventually cross paths.

Among a pool of about sixty tracked ETFs, here are the top ten pairings for the last 250 trading days against the Standard & Poors 500 Spiders (SPY) and iShares MSCI EAFE (Foreign) Index (EFA), respectively, as ranked by the Augmented Dickey-Fuller unit root test.

The pairings presented above each demonstrated a high level of significance (critical values <-3). Don't ask me to speculate on the fundamental underpinnings of each pair result; I'll leave that to you. However, I personally find them occasionally surprising and even amusing! Foreign basket stocks and US Utilities -- who knew?


Jeff Pietsch CFA, Esq said...

That was fast... thank you for your emails! By all means, please post there here!

If you would like to learn more about cointegration, I suggest Carol Alexander's "Market Models" text, which includes MS Excel models. I will add the book to my Amazon links below.

Jeff Pietsch CFA, Esq said...
This comment has been removed by the author.
Jeff Pietsch CFA, Esq said...

A second comment on this article while I'm logged in here. If you are serious about looking into this, look at the ADF over longer time periods (at least 3 years). Lastly, with the explosion of ETFs available these days, I'm updating my tracking list, so there are undoubtedly better pairs out there. Individual stocks can be used as well... it would just take an enormous level of computing power to complete analysis on a meaningful number of stocks.

Christophe said...

Could you share details of the calculation ? I was not able to reproduce your results. I used log price series, adjusted for dividends, for the last 250 trading days and obtained higher critical values (in the -2.5 / -2 range), even while trying various lags in the ADF test.

Jeff Pietsch CFA, Esq said...

Christophe, could you be more specific, maybe focus on one pair?

Christophe said...

For example for EFA/XLU, I get a t-stat of -2.21 at lag 10.

I am using ADF test (with no constant) on the residuals of the EFA/XLU regression.

It could very well be an error on my part. Thanks for your comments.