I did some cointegration tests on a 50 stocks sample, whith which I created pairs. I used the Engle-Granger approach and the Johansen approach, both of them gave me a high percentage of cointegrated pairs(45% and 96%). I'm fairly sure about the way I did the Engle-Granger test (OLS->get residual and test for stationnarity with ADF test), but a bit less for the Johansen test, which I modeled after an example in Jansen's book about machine learning. In the same book Jansen found less than 5 % of pairs were cointegrated.

Does anyone know how rare is cointegration ?

  • 3
    $\begingroup$ Extremely rare, in fact I would argue it only exists for extremely highly related stocks (for example a local listing and an ADR, different share classes of the same stock, two ETFs tracking the same or highly related index etc). Of course a statistical test like Engle-Granger may return a positive result, but that is highly sample dependent. I would bet that many pairs which pass the cointegration test in one period do not pass it in the previous or subsequent non-overlapping period. $\endgroup$ May 26, 2022 at 10:18
  • $\begingroup$ Thank you for your answer. It's good to know how rare it is, my results didn't feel right. Do you know of a way to confirm cointegration ? I tried to do it by doing the two tests and only accepting results confirmed by both. But there are still too many pairs. How can I filter ? $\endgroup$
    – Alec Ric
    May 26, 2022 at 10:54


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.