In the pair trading lesson the authors try to give an example of cointegration without covariance and end up taking two time series one of which is gaussian and the other a square wave and use statsmodel's coint to check for cointegration and claim due to a low pvalue that there is conintgration.
However, this is just plain wrong because the coint function work assuming that the time series are "Variables in y0 and y1 are
assumed to be integrated of order 1, I(1)." as can be seen in the documentation. Sure enough if one takes the cumsum of the Gausian the p-value is quite high.
I find it odd that lectures have such fundamental mistakes and was trying to see how to contact the authors but do not see any feedback place.
