APBRmetrics

DSMok1 wrote:If you use it for multiple years, what is the R^2 of the current regression on multi-year RAPM? (Like I did for ASPM?).

Don't know about multiyear. For single seasons R^2 is around 0.35 for players with total possessions>1000

Also, you might look into a team adjustment? For defense, that can be pretty significant, because box-score stats without a team adjustment tell you very, very little about a player's impact.

Yeah that's probably the next step. Even though it might be almost unnecessary for seasons after '00, because RAPM will (mostly) take care of this; it will definitely improve the accuracy of the ratings before '01

J.E. wrote:

DSMok1 wrote:If you use it for multiple years, what is the R^2 of the current regression on multi-year RAPM? (Like I did for ASPM?).

Don't know about multiyear. For single seasons R^2 is around 0.35 for players with total possessions>1000

Also, you might look into a team adjustment? For defense, that can be pretty significant, because box-score stats without a team adjustment tell you very, very little about a player's impact.

Yeah that's probably the next step. Even though it might be almost unnecessary for seasons after '00, because RAPM will (mostly) take care of this; it will definitely improve the accuracy of the ratings before '01

That R^2 is pretty low. Do you have it split offense vs. defense? I'd bet defense is down around 0.10 and offense is up closer to 0.50.

Are you going to use previous years to inform, again? If you do, please, please derive and use an aging curve. I think the existing RAPM, as currently constructed, has the older players biased high pretty significantly (at the expense of the young players).

J.E. wrote: First of all, since this is not RAPM there's a higher chance that we see players on bad teams rated well, and vice versa.

You stated this earlier--if you scale the results so the team boxscore predictions have to sum to actual efficiency differential (or some thing of that nature) via a team adjustment, you might fix that issue. I'd at least look at that.

And of course
Pts/Poss * Poss/G = Pts/G

DSMok1 wrote:Do you have it split offense vs. defense? I'd bet defense is down around 0.10 and offense is up closer to 0.50.

Good estimation. That's pretty much how they are

You stated this earlier--if you scale the results so the team boxscore predictions have to sum to actual efficiency differential (or some thing of that nature) via a team adjustment, you might fix that issue

My idea was to add
(player minutes)/(team minutes)*(Team OffEff-LeagueOffEff) and
(player minutes)/(team minutes)*(Team DefEff-LeagueDefEff)
then let the algorithm figure out the weights in relation to everything else

I wonder if adding an age/experience factor would improve the rating.

KAN wrote:I wonder if adding an age/experience factor would improve the rating.

Experience is actually in there. Given everything else, it has a negative impact on the offensive rating, and a smaller positive impact on the defensive rating

Another idea I had, which would be a little more laborious to try, is to further break down the assist and block stats into type of assist or block. I know 82games.com has assists broken down into 3-pt assists, jump assists, close assists and dunk assists and the blocks broken down into jump, close or dunk blocks. It also has turnovers broken down into ball-handling, passing and offensive foul turnovers.

Just a thought; but, it might improve the precision of the formula.

KAN wrote:Another idea I had, which would be a little more laborious to try, is to further break down the assist and block stats into type of assist or block. I know 82games.com has assists broken down into 3-pt assists, jump assists, close assists and dunk assists and the blocks broken down into jump, close or dunk blocks. It also has turnovers broken down into ball-handling, passing and offensive foul turnovers.

Just a thought; but, it might improve the precision of the formula.

All of that stuff is definitely coming using playbyplay data from bbr. Right now, I'm just limiting the model to everything available in the 90s to actually be able to rate players of the 90s. Once I feel I've come far enough with this approach I'll start with more detailed player data

J.E., the average major boxscore stats of players with overall ratings of 0, +2, -2, etc. would still be of interest when the timing is right and if it is fairly convenient to calculate. Thanks if you do them.

Alright, I pretty much finished this over the last couple of days, and I have some out-of-sample prediction results that should give people an idea on how to use/interpret those ratings, which I'll publish on my site in the near future.

The contestants were
-Dummy (every player is a 0)
-standard RAPM
-RAPM informed RAPM (RiRAPM)
-new BoxScore metric (BOX)
-BoxScore informed RAPM (BiRAPM)
-Average of RiRAPM and BOX (0.5*(RiRAPM+BOX))

Used as out-of-sample-data were the 07, 08, 09, 10 and 12 seasons, including playoffs. Possessions which included never seen before players (rookies) were disgarded. Offensive efficiency for both teams for every possession was forecasted, instead of point differential (that's how I usually do things, just saying)

Here are the results, from worst to best:
5. Dummy
4. Standard RAPM
3. BOX, RiRAPM (almost equal)
2. BiRAPM
1. 0.5*(RiRAPM+BOX)

I guess RiRAPM+BOX outperforms BiRAPM because, in a way, it uses more seasons. I guess the best thing would be to create a RAPM that is influenced by both BOX and older RAPM (which itself will also be informed by BOX and older RAPM, etc.)

It is exciting to see that BiRAM outperforms RiRAPM. That means progress. Excited to see the results.

Look forward to seeing the new metric data, alongside the older metric data if possible.

Here are the weights I found for offense and defense. Everything's scaled to "influence of height on offense"

height 1 0.8931422196
exp -0.065287672 0.0537292936
GS -0.0185391812 0.1017509227
MP -0.04221383 0.1703430321
FGM 1.0019454696 -0.3614711184
FGA -0.7713214889 -0.0802363173
FG% 0.2848825566 -0.0243612852
3FG 0.2288358653 -0.0252052757
3FGA 0.0349326238 -0.0660744428
3-% 0.3220324436 0.1719594885
FT 0.3366949904 0.1357251166
FTA 0.0736131376 0.0098560925
FT% -0.0252195806 -0.8179984551
OREB 0.296998827 -0.0366778245
DREB -0.1021371556 0.4586301605
ASS 0.4790862015 -0.0081967213
ST 0.0581352101 0.3665636712
BLO 0.0205847853 0.2253025491
TO -0.4038566074 -0.0454896575
FOULS -0.2794180757 0.016693846

I divided the boxscore totals of each player first by (team pace adjusted) minutes. League average is subtracted from a player's per minute totals, the result is being weighed by the listed weight. So if a player had more than average amount of statistic X, which also had a positive weight, that's good for his rating. And
having less than average amount of statistic X, which had a positive weight, is bad for his rating. And
having less than average amount of statistic X, which had a negative weight, is good for his rating.

When I use this new BoxScore metric to build a better prior-informed RAPM, tests seem to suggest that I should use
0.85*(0.65*last_years_RAPM+0.35*last_years_box_rating)
as a prior to compute next seasons RAPM (which then again gets combined with the BoxScore rating to form a prior for the next season, etc)

Height by far the most important indicator for defensive impact, then defensive rebounds, steals and blocks. FT% the runaway most inverse indicator of defensive impact followed by FGM.

What is the average value of a FGM (around 2.18?) and then the break-even FG% point for all FGAs according to this? Is it around 37%?

Interesting about the 2 to 1 RAPM to boxscore weights for the better prior-informed RAPM.