APBRmetrics

If you mainly or only cared about the accuracy of the comparative rpm estimates of the 20-50-100-200 "best" players, is there something you could do to how you minimize the estimated errors or use out of sample testing or something to make the comparative rankings (rather than overall accuracy of these estimates) of just these players more theoretically accurate than found in normal, full rpm runs? I'm guessing yes but am asking the knowledgable. I am guessing this hasn't been done and is fairly unlikely to be done. I think it could be worthwhile in a multibillion dollar industry where history can hinge on a basket or two.

Crow wrote:If you mainly or only cared about the accuracy of the comparative rpm estimates of the 20-50-100-200 "best" players, is there something you could do to how you minimize the estimated errors or use out of sample testing or something to make the comparative rankings (rather than overall accuracy of these estimates) of just these players more theoretically accurate than found in normal, full rpm runs? I'm guessing yes but am asking the knowledgable. I am guessing this hasn't been done and is fairly unlikely to be done. I think it could be worthwhile in a multibillion dollar industry where history can hinge on a basket or two.

If you do your ASPM like DSMok1 and others do it, which is to find coefficients for BoxScore stats so that you minimize the (squared) difference between your SPM and (multi)season (R)APM then you can obviously choose which players to include in the sample and you could choose to run the regression for only, say, the top 50 RAPM guys. The question is whether that will lead to better out-of-sample results (for those 50, or in general). I'm guessing it doesn't, because larger sample size is almost always better. As my professor used to say "there's no data like more data"
Those who run SPM that way should be giving more weight to those players with more minutes, though (because their RAPM estimate is more accurate and their BoxScore stats are more reliable). Since RPM (most likely) correlates to total minutes you're thus giving higher weight to the "top 50/100/..", in a way

Thanks for the reply.