APBRmetrics

Up to him of course.

I very rarely have any interest going back more than 10 years except to track a team trend.

Aren't the BPM's reported on bbref (other than the 2014 and 2015 values) in-sample values? I think this will greatly skew any attempts to blend BPM with other metrics, compare its performance to other metrics, etc.

talkingpractice wrote:Aren't the BPM's reported on bbref (other than the 2014 and 2015 values) in-sample values? I think this will greatly skew any attempts to blend BPM with other metrics, compare its performance to other metrics, etc.

Neil doesn't compare metrics to RAPM, he compares to the subsequent season's performances, testing predictive ability. Also, he runs it back to the 70s, where the RAPM basis only ran back to 2000.

I can't improve much upon BPM without touching RAPM and I don't want to touch any kind of RAPM since there's no data pre-1990. 5% is all I could. One way to improve it would be developing an SPM from 1996-2014 RAPM without age adjustment. However there isn't that kind of data. Perhaps, I should use 14-year RAPM too to develop a metric.

Congrats DSMok1 for such a good metric. However I will improve it without touching RAPM one way or another

Edit: 20% increase in prediction power is the cap for me.

permaximum wrote:I can't improve much upon BPM without touching RAPM and I don't want to touch any kind of RAPM since there's no data pre-1990. 5% is all I could. One way to improve it would be developing an SPM from 1996-2014 RAPM without age adjustment. However there isn't that kind of data. Perhaps, I should use 14-year RAPM too to develop a metric.

Congrats DSMok1 for such a good metric. However I will improve it without touching RAPM one way or another

Edit: 20% increase in prediction power is the cap for me.

Let us know how you do it, and how well it works!

BPM doesn't include a regression to the mean component, currently, but would need such to maximize predictive power, similar to the other box score stats. RAPM/xRAPM/RPM has regression to the mean built in already.

I don't have much time to develop a SPM from ground atm so that I'm using pure box scores (normal and advanced) to come with anything for the blend (no RAPM for obvious reasons) to help improve BPM's prediction power. However it's been a quite challenge since current public metrics already cover much of the box score info.

Currently, experimenting with weighted ridge, elastic-net, lasso regressions and cv methods etc. anything to improve prediction power. (Carefully optimized elastic-net is the winner atm). Gotta be careful about overfitting in the same time.

I'll try to come with a new metric again for the blend tomorrow optimized at targeting BPM's weaknesses and I'll get WP data too. Perhaps WP has some info BPM misses.

No surprise... WP is a dead-end. The worst metric in fact. Even worse than PER.

I need to include players that played in 2 teams in one season for more precise calculations. B-R's player season finder doesn't really help with "TOT" so I'll think something else. Not sure how much a difference would those players make...

talkingpractice wrote:Aren't the BPM's reported on bbref (other than the 2014 and 2015 values) in-sample values? I think this will greatly skew any attempts to blend BPM with other metrics, compare its performance to other metrics, etc.

I fear that's the case here... Because - to my surprise - WS beats BPM in 2014-15 by a quite margin so far. However it's still too early in the season to come to exact conclusions.

Did you see this? https://mobile.twitter.com/NBAcouchside ... 052352?p=v
Is it of interest / use to you?

permaximum wrote:

talkingpractice wrote:Aren't the BPM's reported on bbref (other than the 2014 and 2015 values) in-sample values? I think this will greatly skew any attempts to blend BPM with other metrics, compare its performance to other metrics, etc.

I fear that's the case here... Because - to my surprise - WS beats BPM in 2014-15 by a quite margin so far. However it's still too early in the season to come to exact conclusions.

Beats BPM this season so far? Please explain what/how you're measuring.... Explaining lineup performance?

I need to give BPM's rights back. It looks I made a copy-paste error in the code before. BPM is the one that's noticably better than WS in 2015. Now I have further results for out-sample prediction performance of BPM and WS.

(C stands for centerization at 0 for league total margin which improves in-sample -crossvalidation- and out-sample prediction) I especially used 01-14 data to come with min CV error for both CDBPM+COPBM AND CDWS+COWS. Then predicted 86-00 with those models. RMSE is considerably lower for BPM model which means BPM is a lot better.

Then I used 86-14 data to come with min CV error. Then predicted 2015 season. BPM is better again. Actually RMSE of team margin stands at 20.93534 for BPM while it's 23.6831 for WS.

@Crow

Thanks. I haven't seen it before. I don't think there's much can do to improve upon BPM with 14 year RAPM data but I'll give it a go one day.

This is what I've been saying. I can't yet speak to the newer metrics people have sent me (btw, thanks to all who sent them in!), but plus/minus-based regressions consistently kill WS, PER, and WP.

It makes sense -- those other metrics try to logically intuit their weights, but end up "baseball-ifying" the numbers in a way that doesn't actually hold up when applied to real basketball. By contrast, the weights for something like BPM are derived from what tends to correlate with empirical on-court impact, as measured by plus/minus. So of course BPM is going to perform better out of sample. It's the difference between what works in theory and what works in practice.

Neil, can you give us a 'last call' on these things? I am painstakingly re-creating eWins year by year into the past, and I'd like to get you all that I can muster in the allotted time.

FWIW, eWins adjusts a bunch of parameters and weights in order to force team eWins to match up with pythagorean wins from MOV. So players' stats are indirectly a function of team plus-minus.

Some parameters naturally stay within certain limits. For others, I maintain a floor and/or a ceiling. That's why I wondered whether the Ast*Reb coefficient in BPM should have such a ceiling, so that DReb should not become negative. Then the Moses Deficit might not have to occur, with little loss to overall metric performance.

Mike G wrote:Neil, can you give us a 'last call' on these things? I am painstakingly re-creating eWins year by year into the past, and I'd like to get you all that I can muster in the allotted time.

I was going to start working on this today, so anything last-minute you can add now would be appreciated.

Daniel, do you have the R² for BPM on the 14yr RAPM just for the players with 2270 or more minutes (should be 714 players)? Or can you calculate those? I'm working on a "new" way to interpret boxscore-data while using that subset of players and would like to know whether pursuing that avenue is worth it.