# Cointegration test results differ for different test specifications (drift, trend, etc.)

I am testing whether stock indices are cointegrated. I have two series representing different indices. Both have been tested and are I(1).

When I apply the Johansen test (on eViews), I choose the option for the output to summarise all five assumptions. Under four of the assumptions there is no cointegration however; however, trace indicates two cointegrating relationships if we assume the data has a linear trend.

Since the data doesn't have any linear trend, can this be ignored as a spurious result? Because to my understanding with two series you may only have at most 1 cointegrating relationship?

You should ideally choose the test specification based on subject-matter knowledge before looking at the results. Or you could test which specification is the right one, but that is pretty tricky and may lead to finding cointegration more often than it should be (as you allow the test specification to adapt to the behaviour of the series too well).

One may assume a stock price index to follow a random walk with or without a drift. A drift is justified from the financial theory point of view as a compensation for undertaking risk (compare the "risky" investment in stocks to a "safe" investment in U.S. government bonds).

• If you assume upfront that the series does not have a linear trend (i.e. a drift), you should just neglect the test result for that case.
• If you assume the series has a linear trend (i.e. a drift), you have to deal with the test result for that case.

How come the unit root tests suggest integratedness while the cointegration test suggests no integratedness? This conflict is of course confusing and could not hold in population. Nevertheless, it may appear in finite samples where the unit root test is not powerful enough to reject a unit root while the cointegration test has enough power to reject lower orders of cointegration.

You could either try alternative tests looking for which case receives more support or just assume what you believe to be right and proceed.

• Thanks for your input Richard. You answered most of the question but what about the cointegration tests showing two cointegrating relationships between two series. Would that have to be considered spurious given that methodology is correct? Or simply because the assumptions selected for the test do not match the data. Jul 23 '16 at 15:02
• They probably do not show $2$ but $>1$, strictly speaking. That happens when both series are stationary. Then there is an infinite number of linear stationary combinations (actually, any linear combination is stationary). If that is hard to believe, probably this is a problem of a particular sample where the behaviour incidentally looks stationary while normally it wouldn't. Does that answer the question? Jul 23 '16 at 15:06