Project: Predicting wins based on Box Score totals

[Sacré d'Jeu]

Hello everyone,

I'm a student at University of Ghent, Belgium. We were assigned to do a project for the course Machine Learning and I had the idea of doing something with NBA statistics.
More specifically, had the idea of trying to predict which team will win the game, based on their average box-score totals.

Can anyone help me to get data, containing box-score totals of every game? (like the one on NBstuffer, but for free )
Alternatively, I think I can use the data on basketballvalue.com and work with 5pl matchups instead of teams.

And do you find it a good idea? Any suggestions or other ideas?

[v-zero]

I would suggest that a team based model combined with a player based model and some machine learning techniques could produce some interesting results. I have the last ten years of so of box scores (game-by-game) available in a Pickled Python Dictionary format if that interests you?

[Sacré d'Jeu]

I think I'm gonna start with a model, based on team statistics, and then later expand this model with player stats.

The box scores would be very welcome. I have no experience with Python, but fortunately, I have a colleague students that might be able to convert them in something I can use.
Most likely, I will program in Matlab, as this is the software we also use in labsessions of the course.

[v-zero]

Ok, hopefully they'll be simple to turn into a suitable .mat file. Drop me your email address by PM and I'll send them to you.

Will you be intending to share your eventual work/paper with us here at APBR?

[Eternal]

Yes, one interesting approach would be to fit a team level power ranking model, then fit another model to predict offensive and defensive strength as a function of other statistical parameters.

-Chris

[Sacré d'Jeu]

Will you be intending to share your eventual work/paper with us here at APBR?

I will update here how things are going, and eventually post the resulting paper.

Yes, one interesting approach would be to fit a team level power ranking model, then fit another model to predict offensive and defensive strength as a function of other statistical parameters.

I'm not sure you mean... do you suggest one or two projects?

I think I'll do the following:
- find how much each statistic and parameters (like home/road team) influence the outcome of a game
- predict the outcome of a game based on their average statistics/parameter

[Eternal]

Power ratings typically consist of offensive and defensive ratings. These, in turn, could be estimated from such parameters as rebounds, steals, turnovers, 2-point shots, 2-point efficiency, etc.

-Chris