Sharing code for RAPM-style player rankings

[Crow]

When I talked about duos above I was just adding the individuals myself. I temporarily forgot about the 2 player data set. I'll look at that a bit later.

[Crow]

Looking at ridge 2 player vs ridge 1 player estimates for OKC, GSW and SAS they seem to tell notably different story for each team, with the sum of the 3 parts of ridge 2 only being about half of the sum of the 2 parts of ridge 1 combined. I'll try to better understand what you just said but if you can comment on this in different, more obvious or thorough language I'd appreciate it. I know all of Durant, for example, isn't the same set of data as KD with Russ but with Russ is a high % of all of KD and doesn't this data suggest KD w/o Russ is about as good as KD and Russ together combined, maybe even better? I'll check but I'd be surprised if the raw plus minus data looks that way for OKC, GSW, SAS and other sets of stars. And enet and lasso team results on the heat map confuses me further. GSW sucks there because... Curry is great but most of team minutes aren't Curry? I am sure I need more explanation of the "team effects / coefficient" and how that fits with the different levels of data.

Are you planning to do 5 man estimates? If not, is it soley because of sample sizes? It might still be interesting to see all the interactive terms. Would there still be a team coefficient here?

[mscbuck]

I know all of Durant, for example, isn't the same set of data as KD with Russ but with Russ is a high % of all of KD and doesn't this data suggest KD w/o Russ is about as good as KD and Russ together combined, maybe even better? I'll check but I'd be surprised if the raw plus minus data looks that way for OKC, GSW, SAS and other sets of stars

I won't give away my somewhat work-in-progress method (though you can find it by looking at other-sport analytics, sorry, I can't give away my secrets! ), but it's useful for attacking those type of "within-team"/"within-lineup" questions, Crow. Just looking at my numbers and KD/RW, RW's contribution to his lineups performance is only slightly higher than KD's in lineups where the other is out (aka, KD w/o Russ is about as impactful as Russ w/o KD), but when in lineups together it's Russ's contribution that is higher than Durant's and is closer to his overall "without the other guy" contribution. However, a slight oddity I do see is that KD has an overall higher "replacement" effect than Westbrook across all lineups (aka stick him in any lineup and they'll tend to do better), which may be speaking more to the results that wc3 finds that show KD potentially ahead of RW. Just thought I'd throw some additional information!

[Crow]

Thanks for at least indicating that this with / without issue is something you pursue. I might look for the approach you hint at and see if there is some new way to analyze beyond what I have done in past using with / without splits from 1-5 player raw and adjusted plus minus and associated box score data.

[wc3]

wclark3.github.io/basketball

Just updated the rankings to include last night's games. Plan is to update once weekly (meant to do it last night but the data source changed format over the ASG break and I had to re-write my scraper). I won't post here every time, just thought it would be relevant to say it once though.

[Crow]

wc3, any thoughts to share about the different value messages at the 1,2,3 player levels? Any plans to release 5man data?

[cemitten]

WC3, is the basketball portion of your site down now or permanently?