Starting line-up

[Hiroaki-K4]

Hi, there.
I would like to see if I can improve the accuracy by adding the starting line-up stats information. If anyone has already tried this, I'd like to know the results. Also, it would be great if you could tell me how you acquired the data.

[nealmick]

I have built models with season averages using team rosters before. The data is available via several api's I have used one called balldontlie.io I also have a branch that has some logic to use last game data of all the players stats here, although i haven't messed with it recently. One of the main issues is injuries and accounting for different rosters, trades and so forth, some tricks are to use the same players as last game for rosters, otherwise historical injury data is somewhat difficult to find. You can kinda cheat and use the players season averages that actually had play time in that game to train, but for future games you don't actually know who will play, which can cause issues if in training you knew magically certain players were benched. Future games you wont have that same retrospective data.

[Hiroaki-K4]

Okay, thank you for sharing information.
Could you improve accuracy by adding player stats?

[nealmick]

No i cant. Pretty much the model uses odds as input feature, so it knows market value. The market is extremely efficient, and accounts for all stats and all information generally available. All information is priced into the market. It is near impossible to build a model that is more accurate then vegas odds, and if you use vegas odds as input, it will be as accurate as the odds and any other added stats dont really provide any value. If you build a model that is not market aware then it will have lower accuracy in general but you can then see improvements as you add more data. Player stats do correlate to game outcomes, and they are pretty much the best data we have, the most important stats are a teams win/loss record and streak, player stats are more specific and offer less impactful information then meta stats or team averages. the top players' 3 point shots have fairly high correlation to game outcomes.

There is also way too little data/games, and the answer answer to who wins a game changes over time, the game changes so that different strategies matter more and the further back you go the less relevant the data becomes to the current state of the game. There are only 10-20k games if you start from around the 2000's depending on your criteria and so forth, which is a relatively small amount. Lastly the game of basketball is fairly random and one team may win sometimes and not other times. Also its possible depending on the data to have the same input for two different games or very similar input and then get different results.... all this combined makes it near impossible to model the game, and the best approach is price discovery making people put money on the line and letting them use all information publicly available, which is casino gambling.

[Hiroaki-K4]

Thank you so much for sharing information.