I could be totally off but when performing a regression a constraint needs to be implemented. I believe there is a flaw in the above analysis. I stumbled upon this possible error as I have been reconstructing "Analysis of sports data by using bivariate Poisson models" by D Karl's et al and applying regression to the data above.

If you were to do this

response = Riffle[hgoals, agoals]
GeneralizedLinearModelFit[{X, response}, ExponentialFamily -> "Poisson"]

You get an error, "The rank of the design matrix 40 is less than the number of terms 41 in the model. The model and results based upon it may contain significant numerical error."

There needs to be a constraint in the regression. Maybe I am incorrect??

Hi Sam!

First and foremost I should admit that none of our group had ever used Mathematica prior to undertaking this project, so were very much learning as we went along. At the outset we did look at those two machine learning based models, but had difficulty deciphering all of what was going on.

And since we wanted to look at club teams rather than international teams, there was much less data to work with. Most significantly we were unable to obtain data on individual players, so building a classifier may have been difficult.

The general model we used is called time-independent Poisson regression (see here) and although perhaps not the optimum model for this task does obviously perform reasonably well.