eminence wrote:OhayoKD wrote:-> For RAPM, A penalty is applied to "outliers" so that they "converge towards zero". In other words, the gaps between players here are suppressed, and not actually indicative of what they would be in the real-world. "Closeness" can only really be gauged here in a relative sense(ex: gap between #1 and #2 vs #2 and #3), an extrapolation like "player a is worth 15% more than player b" doesn't really work.
So I haven't done any large sample APMs/RAPMs recently, but from my understanding I don't think the regularization penalties are a serious concern on a sample of this size (at least looking at players with somewhat similar RAPM numbers, it's probably a little concerning if you're looking at the guys at the top and bottom of the list).
Your optimal regularization value will grow with approximately a parameters/sample size ratio. Very large sample size = relatively small optimal regularization value.
This is a lot of guess, as I don't really know how Cheema did theirs, but on past experience I would guess on a sample this size it is only having a couple tenths of a RAPM point pull at the edges, scaling down to 0 at 0 (duh).
Eg: LeBron's 5.54 might 'really' be a 5.8 (+0.26), while KD's 3.54 might 'really' be a +3.7 (+0.16) and so on.
Good to know!