SideshowBob wrote:RPM is a regression based model that uses SPM as a prior, it is similar to other box&+/- hybrid metrics such as IPV, but it is not a blend in the same manner. This kind of stat needs a considerably larger sample size for stability; I wish ESPN had held out, just for the sake of irrational responses that we'll now see.
RPM (in this case) is meant to say the following:
All else held equal, if [Player X] is in a lineup, the lineup's performance (MOV per 100 possesions) is expected to change by [Rating] per 100 possessions.
It is NOT a catch-all player rater.
-----------------------------------------------------------------------------------------------------------------------
A valid statement I can make based on RPM:
"If I put Steph Curry on any random team, I expect the team performance (MOV), while Steph Curry is on the court, to improve by 8.85 points per 100 possessions, given the the league-wide lineup data we have for the first 3 weeks of the 2016 season."
An invalid statement:
"Kyle Lowry is the 3rd best player in the league."
Sorry to nitpick but I believe the first statement is also invalid. Given that our sample size is so small, the +8.85 estimate
likely has an enormous confidence interval and the 8.85 itself is almost certainly not the true MOV parameter. In an extreme case, it may not even be significantly different from the 0 null (though in all likelihood, it is). I'm also not too familiar with the construction of the regression model but does anyone know if the model contain interaction terms?
Also the generalization here about any random team is somewhat dangerous given the importance of context and roles, which the regression model is unable to capture.