APBRmetrics

http://espn.go.com/nba/statistics/rpm/_/sort/RPM

I am not particular about what gets discussed within a single thread; but some are. So fwiw, here is a separate thread for this angle, if any others want to use it too.

If Zaza ranked 16th highest on RPM estimate is a conversation killer; then instead ask how many of the top 20 are in reviewer's estimation top 20, 30 or 40. If they say they agree that 18 of the top 20 on RPM estimate are in their own view top 30 or 40, then you have laid a broad base of pretty similar evaluation at the top and that should help in rising above 1-2 substantial disagreements or maybe even a few more. Top 20 is a bit better than top 5%. Top 30, top 7%; top 40 within the top 10%. A disagreement over placement in top 20, 30 or 40 is splitting hairs to me.

If you take the entire RPM set of estimates and divide the ratings into a 5 tier system of great, good, average, below average and well below estimated impact, how many of the estimates would be judged by an insider as more than 1 tier away from their own independent estimates? How many of the estimates on a specific insider would be judged as being more than one tier different than the evaluation of a second insider? Some disagreement is to be expected. If a RPM skeptic or critic would actually do this, it would be interesting for the discussion. Take the list and mark it up. Do you disagree by more than 1 tier on 20%, more or less? Don't just cherry pick a few. Mark up the whole list and tell us how many are really "off" by more than one tier. There probably are some bad estimates. Looking at everything can help pick them out. Any method is probably going to have some outlier mistakes. Do you throw the tool away because of a few bad swings? And if you find that you agree within one tier on 70-95%, will you tolerate the tool or use it alongside everything else (which are imperfect estimates too) Who is being unreasonable or strident if you won't do that or tolerate that?

12 pages of RPM estimates. How many of about first 2.5 pages "belong" after 5th or worse? How many after 2.5 thru about 5 belong after 7.5 or worse? How many after 5 thru 7.5 belong on 1st 2.5 or after 10th? How many after 7.5 thru about 10 belong in top 5 or bottom 2? How many in bottom tier belong in top7.5? Add them up.

You can count 1 tier variances too if you want.

On first 2.5 pages I see only about 20 cases that could imo conceivably be better suited for worse than 5th page. So I speculate a minimum 80% agreement within one tier, 20% disagreement. If you count more that, name them.

Myself, on very quick review, have plausible skepticism on no more than 10. I'd guess I'd end up with about 5 if I reviewed further and committed to yes / no guesses of variance of more than 1 tier. So top tier, I am saying I agree within one tier on about 95%. How about you?

# on pages 2.5 - 5 who are below average? I'd say no more than 7-12. I'll go with 7.

# on pages after 5 - 7.5 who "belong in top 2.5 or bottom 2? I'll go with 10.

# on pages after 7.5 - 10 who belong in top 5? I'll go with 7.

# of pages 11-12 who belong in top 7.5? I'll go with 5.


So on quick pass I estimate that I think about 34 RPM estimates are off by more than one tier or about 8%. Or about 92% are within one tier for me, on quick pass. Maybe I was pre-disposed to roughly this level of confidence but that is my finding. It could be less. I was trying to find likely cases. Of course I may have missed / be blind to others. I hope others try it to help assess range of confidence / skepticism. Anyone score over 20% disagreement by more than one tier? Buehler?

I'd go with a blended metric over RPM alone or anything alone.

How many of this top 20 viewtopic.php?f=2&t=8962 would you say are not top 20, 30 or 40? How many of this top 100 are not top 100 or 150? Goal is get close or closer than with less rigorous means. And to think further about the outliers. More right or wrong? Why? In the end, it is about using the numbers, working the cases and not being worked by the numbers or insufficiently working / thinking about the numbers.

Think Zaza's RPM is off or way off? Think RPM is a black box? Please ESPN and RPM authors bring on RPM factors. If you choose not to work any of the dozens of other split opportunities anytime soon, please push the discussion forward by providing the RPM factors at least once a season.

Let's see Zaza's RPM factors. What is RPM claiming but not saying anything explicitly are the tangible specific reasons for his good RPM estimate. It is clear that offensive and defensive splits of RPM are not enough to satisfy the critics. Nothing may be, but give the tool users more to consider.

Any willingness to back off the estimate to 2 decimal places? That breeds over confidence amongst some and contempt from others. Back off to one decimal or nearest .25 or .5? Given past error estimate talk, any one is more reasonable than taking it out to 2 decimals. Any willing to show Tiers or at least discuss the concept?


Lets compare RPM and its factors to 3-5-10 other versions of RAPM and their factors and to 3-5 boxscore metrics and their factor splits. Where are the discrepancies for this individual and most frequently leaguewide? Exact matches might be interesting or aren't necessarily the goal. Perhaps there is some middle ground involving tweaking one or more metric that might be estimated "better" than today's mix though. At least possible. Or call the hunt off and decide to stay this imprecise and in this level of disagreement.

Pachulia's SPM prior is ~0.8 / 1.8, for offense / defense

Per 36m, he averages 12.6/10.4/3.7(AST)/1.7(STL).

He's 7th among big men in AST%, 3rd in STL%, 17th in the league in OReb%

When you look at MIL's +/- from last season http://www.basketball-reference.com/tea ... 15/on-off/ you see that only Middleton, Pachulia and Dudley have a positive NET among the 7 players with most MP played, with Middleton's and Pachulia's NET being rather large. A situation like this rarely occurs. It's more common for e.g. an entire starting unit to have good +/- (like this http://www.basketball-reference.com/tea ... 15/on-off/).

In a situation like Milwaukee's, the positive credit gets shoveled onto only a few players: Khris Middleton, first and foremost; some goes to Pachulia and (less) to Dudley

Bradford Doolittle's recent work uses 6 tiers for the valuable players (A seventh tier for the rest?) Is something in that range (or the 5 I suggested) about as many tiers as may be "needed" or used well / broadly? If you did a cluster analysis on non-rookie free agent salaries would you find 5-7 clusters did a good job of classification? I'd guess so. Does analytics need to conform further to common patterns of organization and precision?