drewlanenga
commented
10 years ago A cool approach to play-by-play analysis:
Blog: http://grantland.com/features/expected-value-possession-nba-analytics/ Paper: http://www.sloansportsconference.com/wp-content/uploads/2014/02/2014_SSAC_Pointwise-Predicting-Points-and-Valuing-Decisions-in-Real-Time.pdf
First, Update
I've tweaked a bunch of stuff, but haven't committed most of it...
I toyed a bit with the idea from @drevell to do some parametric distribution approximation on top of some of the MCMC output. I remembered near the end that the posterior draws for all the beta parameters are (intentionally, and should be) correlated. So drawing independently from approximate parametric distributions won't work...
Also, I added location to the model, improved offensive/defensive ratings, but still am only hitting at about 0.6 on a log-loss score for the BEST seasons. (Near about 0.9 across all seasons.)
Second, Pivot!!
Rather than just iteratively tinkering with the boxscore model, I'm thinking of a possible "pivot" in the model direction.
Since the approach is simulation based, it'd be great it we could do some work on simulating individual game plays. To do that, we'd need to scrape play-by-play data, whereas now we only have boxscore data.
It'd require a bit of a different approach, but I think it could be pretty cool. @drevell: Let me know if you'd be interested in helping at all with any of the new scraping or model building.