APBRmetrics

Neil Paine · Joined: 13 Oct 2005 Posts: 774 Location: Atlanta, GA

Yeah, Efficiency Differential is just Points Scored per 100 Possessions minus Points Allowed per 100 Possessions.

Mike G · Posted: Mon Nov 10, 2008 12:51 pm Post subject:

I am not where my files are, but I'm sure I gave Ginobili reduced minutes. (I used the same round-number estimates as I applied to my own eWins-based predictions.)

Meanwhile, I'm wondering about the way previous seasons' APM data are infused into the estimates. Are rookies and others with fewer seasons based entirely upon their season(s) played? Should their (nonexistent) previous years be accounted as zeroes? Effectively, that's just regressing to the mean.

Instead of weighing previous seasons successively lighter, how about weighing them the same (depending on minutes, error, etc) and deriving for each player a sort of "effectiveness/productivity ratio". Maybe Al Jefferson has great numbers and low +/-, and so one could say he's only half as good as his stats. Bruce Bowen might be twice as good as his stats.

I know I'm stating this indistinctly. But when a player's production rates go down, his APM might be be estimated as proportionately lower.

Dan R. noted in his original study that more experience players were more effective than younger guys, given the same stats. So I guess recent seasons should get a bit more weighting. But for someone in a statistical free-fall (Shaq?), his current stats, adjusted to his 4-year 'E/P' may be a better gauge than a uniform 4-yr sliding scale would be.
_________________
`
36% of all statistics are wrong

Mountain · Joined: 13 Mar 2007 Posts: 1527

I know a model that perfectly fit last season's results isn't assured to do the best in the future

(and maybe there are multiple solutions equally as good?)

but with the elements being discussed (age adjustment, multi-year weighting of some flavor) for whatever "box score based statistical" component(s) that is used (though maybe split offense and defense for clarity and perhaps separate weighting)

and adjusted definitely split into offense and defense for separate weighting

what weights does that last year's "best fit" model use?

As for my divergence from what adjusted +/- projections for next season I was probably trying to make some assumptions about year to year consistency of statistical vs adjusted on the fly and might not have been consistent about how I did it.

I am not a huge fan of TJ Ford but figure he has to be better as team organizer than Tinsley and in line with their talent/ desired run and gun style. But if he dominates the scene (and specifically brings his awful last season adjusted +/-) then my projection may be too high. I think I was counting on him to sort of blend in pretty well to what they had going and same for Nesterovic and the rookies.

As I said before adjusted +/- for defense would benefit from a counterpart / help split and offense from an own offense / team offensive impact split and I think this could be accomplished though others could say better.

Right now in pure adjusted the individual statistical contribution is fully there but in adjusted terms. It would actually be double counted if it is an additional element to statistical in the above model I outlined. If you had adjusted broken out into 4 parts individual and team impact, offense and defense I'd try ditching the legacy statistical information or play further with weighting the two and see what gave the best fit and felt the best.

Ilardi · Joined: 15 May 2008 Posts: 262 Location: Lawrence, KS

In reply to the trenchant points raised by Davis, Mike G, and Mountain (true identity unknown?) -

Yes, I agree: the published 07-08 APM numbers (based on weightings of .125, .125, .125, .2, and 1 from 03-08 ) may not yield the best possible prediction of 08-09 performance. I've actually thought about calculating a long-term average value for each player based on an unweighted composite of 03-08, and then deriving an individualized "growth curve" for each one based on the component (weighted) single-season values over the 03-08 window. The ensuing player year-by-year trajectories could then be projected forward into 08-09, presumably smoothed out some based on aggregate league-wide APM growth curves linking APM to age, yrs, and position.

It'll be a big project, but should prove worthwhile . . .

Mountain · Joined: 13 Mar 2007 Posts: 1527

Lots of weight sets possible.

One based on just 2-3 years with the prior yr(s) given a substantial weight might be good on a future run.

Mike G · Posted: Tue Nov 11, 2008 7:21 am Post subject:

OK, here is the minutes-distribution that predicted 60 wins for the Spurs; and an alternative that takes 600 minutes from Ginobili, allotting 200 each to Udoka, Mason, and Finley.

Neil Paine · Joined: 13 Oct 2005 Posts: 774 Location: Atlanta, GA

I don't think that's right, Mike... The first distribution of minutes yields a weighted average efficiency differential of +1.71, and the second gives you +1.45; according to the formula I posted earlier in the thread, those differentials should give you 45.3 and 44.7 wins, respectively. If Ginobili doesn't play at all, Udoka plays 1867 minutes, Mason 2067, and Finley 2467, the efficiency differential is +0.83, good for 43.1 wins.

Brian M · Joined: 25 Nov 2006 Posts: 40

I took a look at correlating the adjusted +/- data to 2007-08 PER for players with >= 1000 minutes played. (I suppose it would be more appropriate to take a weighted average PER over several seasons mirroring the apm data, but alas.)

correlations of PER with:
total apm: r = .50
off apm: r = .58
def apm: r = .05

Regressing total apm onto PER:

apm = .4954 * PER - 7.5031

So a rough but simple rule of thumb appears to be apm ~= (PER-15)/2, though the estimate is rather noisy (std dev of residuals = 3.4) because PER does not capture any variation in defensive apm.

The correlations between ORTG and off apm (r = .48 ) and DRTG and def apm (-.49) are pretty solid. Interestingly, usage rate correlates with offensive apm (r = .44) but not ORTG (r = .02), consistent with the hypothesis that usage has offensive value. (Though alternatively, perhaps players get higher usage in virtue of whatever they do offensively above and beyond what shows up in ORTG.) Doing a partial correlation to control for differences in offensive apm, a negative correlation between ORTG and usage emerges (r = -.24, p = .0001)

Taking the average of zscore(ORTG) and -zscore(DRTG) gives correlation with apm r = .5. apm correlates with win shares slightly but perhaps not significantly better than PER (r = .55)

Mountain · Joined: 13 Mar 2007 Posts: 1527

Good stuff Brian.

Would you have interest in checking correlation of say 1 or 1.5 3 pointers made per game and offensive APM or 5+ assists per 48 and offensive APM?

Is usage a good thing or is usage at a respectable or better efficiency a good thing with most high usage players meeting that criteria? If you divided high usage players (at whatever level) into high/low or high/middle/low efficiency would the positive APM correlation be largely unaffected for the distribution or greatly favor the efficient or disappear for the inefficient?

Anybody apply adjusted +/- methodology to say football receiving corps vs the secondary to see if high usage receivers improve performance of other receivers above when the high usage guy is not there? Or apply it to soccer or volleyball or any of a number of other sports to look at the same question?

Mike G · Posted: Wed Nov 12, 2008 4:58 am Post subject:

**cherokee_ACB** · Joined: 22 Mar 2006 Posts: 157

Mike G · Posted: Wed Nov 12, 2008 6:38 am Post subject:

Neil Paine · Joined: 13 Oct 2005 Posts: 774 Location: Atlanta, GA

Brian M · Joined: 25 Nov 2006 Posts: 40

Mike G · Posted: Wed Nov 12, 2008 10:53 am Post subject: