OhayoKD wrote:f4p wrote:OhayoKD wrote:Seems about right, yeah. Though maybe someone whose good at this stuff like unibro can check we're interpreting this properly
so this seems to get at the point i was talking about in our other back and forth about all of these metrics:...but is still going to be a struggle where i think the error bars (i.e. who actually gets the credit from the lineups) probably exceed the differences we are talking about...
setting aside PER for a second, which isn't even trying to predict anything, the error on EPM is 2.48 and the error on BPM is 2.7 and for WS48 is 2.85. the difference in their errors is an order of magnitude smaller than the error. that's a difference (from EPM) of about 0.6 wins for BPM and 1.0 wins for WS48, presumably against a background guess of 41 wins for the average team. and comparing to RPM basically eliminates any difference with BPM and makes the difference with WS48 about 0.7 wins. these are small differences. EPM isn't God talking to us on a mountain, except instead of the 10 commandments, he's giving us his top 10. and BPM/WS48 isn't a shirtless drunk guy yelling at the end of the bar. even PER manages to trip and fall, without even trying or being the same thing, into only about a 2 win difference with EPM and 1.6 win difference with RPM. the overlap of the predicted value for individual players on these things is going to be pretty substantial and leave us well within the zone of interpretation of what matters more for each player. and of course, the RAPM/PIPM stuff is inbetween the BPM and WS48 errors.
I'm really not smart enough to interpret the 20% thing you threw out,
what 20% thing? you posted the article, not me. i'm just reading their own numbers which say they are at best improving from 2.8 to 2.5 on prediction error, and only with one specific metric is the difference even that large. also, from that nyloncalculus Ben Taylor article, which is quite similar, we have the following chart:
so within a margin that anyone would give a tinker's damn about, WS (i assume he meant WS48 or maybe his minutes adjustment makes WS and WS48 the same?) and BBR BPM are practically identical to PIPM and backpicks BPM (that's 2 of the 3 BBRef stats). The Y-axis starts at 2.5 so if it actually started at 0 (i.e. zero error), the differences would look pretty damn small. even at the lowest continuity part of the graph, which certainly does not apply to all teams, the difference is not very large, and then then only seems to apply to PIPM, with the gap b/t PIPM and other metrics quickly reducing as continuity goes to 65% (APM and AuPM from previous graphs would look very similar to the other 3 non-PIPM stats). and some of the unspeakably horrible stats seem to actually predict slightly better (still no real difference) if teams have high continuity. y'all are the ones posting the articles, i'm just taking the info from them. a 0.2 difference in predicted net rating is a difference of 0.5 wins. over a whole season. the statement from some seems to be that one prediction is the greatest thing ever and the other is so stupid no one should ever consider it. the evidence seems underwhelming on this front.
this doesn't seem like so I'd rather let you and unibro duke it out(10 wins would be a pretty big gap, 1.2 wins i guess not though maybe that snowballs into bigger gaps in players).
if unibro has articles with 10-20 win differences, then he should post them. perhaps we will see something the other articles haven't mentioned. Ben Taylor and dunksandthrees seem to have much different numbers and differences and don't seem like unreputable sources. and, either way, weren't chosen by me.
I think I addressed "its not meant to predict winning" in the other thread though and I think some of that stuff applies even if the stats were equal or edging to the box-stuff. If you wanna continue that convo we should probably do it there
i will read the other responses at some point and maybe respond here.