>do we calculate the bayes factor by using the probability of winning *for the winning team?*
in a sense yet. We have to compute the probability of seeing the observed result (e.g. a win for reds if that is the result) given the model is 'correct'. That is what is meant by P(D | M) where D is the observed result and M is the model M is correct.
So if the model M says there is a 0.9 chance of reds winning this particular match and we observe that reds do actually win that match then P(D | M) = 0.9.
But you have made a mistake when you say:
> and .99 and .01 for M2
The .01 should be .91 because for match 5 model M2 says there is a 0.91 probability that blues win and the result is a blued win.