I'm working on a way to simulate NBA games where I simplifying all the conditional probabilities by making Markov chain assumption that the probability of each event is conditional only on the previous event. This seems like a really nice compromise between disregarding conditional probabilities altogether and overfitting probability estimates that are conditional on the previous 100 events.
Also, I think very few basketball fans think the event probabilities depend on much more than the previous event.
Well, also conditioned on the players on the court. And the coach?
I'm going to use a baysian approach to estimate the conditional probabilities, but I don't think the prior is too important because there are so many events if you only condition on the previous event.
So far, all I've made so far is a giant database of every possession since 2017 and every detail about the possession. I will email it to you if you ask.