The percentages are basically the score probabilities. My model, for instance, calculates that Fulham has a 45% chance to win against Chelsea, but the probable scorelines are quite evenly distributed, the biggest being 2 -1 for Fulham at 10% (although it also predicts 1 - 1 at 10%). Rather unexpected I must say, I guess Chelsea's performance against Man City really weighed them down here.
And yes, those are using Poisson for now. Though will probably be looking to fit negative binomial distribution to my model as well, just to try it out, that is if I can somehow manage to find a way to roughly estimate the goal conversion rate for each team within a particular match and if I find the time and energy to do it (highly unlikely nearing the end of the year!). For the time being, esp as I'm only starting with this score prediction malarkey, I'm quite content with using Poisson and will first look at whether introducing arbitrary weighting factors and similar small alterations to the equation can induce enough bias in the model to overcome the shortcomings of the distribution (instead of using long winded mind boggling algorithms ).
Nope, no can do ZIP model with SPSS (the same with negative binomial models as far as I know), I can use other softwares for that, but in the meantime will be relying on good' ol Poisson regression. You're planning to do ZIP with R?