APBRmetrics

v-zero wrote:Like I said, single season NPI RAPM is crap.

permaximum wrote:

v-zero wrote:Like I said, single season NPI RAPM is crap.

Then how are we going to judge player performances season by season?

On the contrary, single season non prior informed RAPM is the best for my needs. I want to analyse players' seasonal performances.

permaximum wrote:Then how are we going to judge player performances season by season?

On the contrary, single season non prior informed RAPM is the best for my needs. I want to analyse players' seasonal performances.

EvanZ wrote:So I'm doing 1-yr RAPM with no prior and finding something very, very strange (or maybe not considering the post above):

https://docs.google.com/spreadsheet/pub ... utput=html

No, not that Amir Johnson is near the top. Or Chris Paul or Tim Duncan or Kevin Durant.

What is strange is that LeBron is at #363 with a rating of -1.70.

I'm calculating RAPM exactly the way I've done it in the past (using cv.glmnet to find lambda that gives minimum cv-error). The lambda I'm getting is ~100. This is puzzling, to say the least.

v-zero, have you found something similar when you do 1-yr non-prior-informed ridge regression?

v-zero wrote:This is basically a data dump as a result of incomplete work I'm doing - I know people are missing xRAPM and IPV (and most largely the lack of a good defensive metric) so hopefully this will help. It's built from 2000-01 onwards, with the playoffs, and with each year's ratings as the prior for the next. There is no box-score information involved, and new players to the league receive a prior according to their percentage of possible possessions played - so in 2000-01 everybody got a prior based on their percentage playing time, basically. Lambdas were found via cross-validation until it was clear that lambda settled (i.e. that the ratio of useful information contained in the prior to useful information contained in the in-season measurements became constant, which more or less happened in 2003-04). The ratings are such that possession weighted offensive ratings sum to zero, and possession weighted defensive ratings sum to zero, so that these can be interpreted with zero as the origin for for defence and offence.

Single Year Prior Informed RAPM:
https://docs.google.com/spreadsheet/ccc ... sp=sharing

Single Year Non Prior Informed RAPM:
https://docs.google.com/spreadsheet/ccc ... sp=sharing

N.B. I'm not suggesting anybody have too much faith in these, given that they aren't my end goal, but they should be as good as the estimates JE made with RiRAPM.


EDIT:

These have been discussed somewhat at gamefaqs and possibly elsewhere, I don't post there so I'll post here: Don't trust these ratings as prescriptive. They're a work in progress. Specifically intelligently decaying 'old' information (data from prior years) and weighting priors by sample size (in a basic sense) is yet to be properly done, so when you see KG and Amir ridiculously high on D you should assume the number is too high, but that they're pretty excellent. As for guards being underrated on D: size matters, guards are nothing like as useful as bigs on the interior, and since shooting% is roughly inversely proportional to distance it's fair to assume that great interior D beats great perimeter D. However, guys like Battier whose awesome perimeter D doesn't show up in the box score will show up here.

EDIT 30/03/2013:

Updated with most up to date pbp data and a vast improvement on the priors.