Right now I'm using orating / drating / a proxy parameter for tempo (which I think needs improvement) and a variable for home court advantage (found this to have only a smallish effect +/- 2 points maybe).
I ran a regression using essentially the kitchen sink of possible variables and these were the ones I found that were pretty significant on a teams score.
I re-ran the regression using these core variables and took the coefficients and weight them accordingly for each match up. For orating/drating I have the ability to use a lot of "different" versions of these variables ie, BTB / over the last 5 games / year-to-date, etc. Right now I've been purely using YTD in the predictions that I have been posting. I have been toying around with using more "specific" inputs with some success I think.
I kind of am waiting to see when this model will actually fail. Right now, if you say picking "over/unders" is distributed like a binomial variable, and the probability of success is 50% the odds of getting AT LEAST 21 correct out of 32 is 5.5%
I would like to think that, that is something to be happy about, but right now I need a larger sample size.