Reconstructing Box Plus/Minus

Home for all your discussion of basketball statistical analysis.
Crow
Posts: 6250
Joined: Thu Apr 14, 2011 11:10 pm

Re: Reconstructing Box Plus/Minus

Post by Crow » Fri Jun 14, 2019 12:39 am

Can you give before and after example for:

one specific PG and one specific other position player of how much credit is in the end actually being given for assist behavior

one specific PG and one specific other position player with the same levels of assists

and / or for the average PG and the average for that other position?

vzografos
Posts: 24
Joined: Thu Sep 06, 2018 10:42 am

Re: Reconstructing Box Plus/Minus

Post by vzografos » Fri Jun 14, 2019 8:16 am

DSMok1 wrote:
Thu Jun 13, 2019 4:47 pm

Gradually working on this project!

Yes, the idea is to maintain the existing general concept of BPM (i.e. historic applicability, general structure) and significantly improve the handling of outliers.

Thus far I have focused on the linear version of BPM, currently called GmBPM. It should be very stable and should handle outlier numbers very well.

Then I intend to build upon that framework and add nonlinear terms as appropriate to help handle nuances while hopefully not destroying applicability to outlier values (like the existing BPM did).
Maybe you already mentioned this, but how does your new equation look like (GmBPM)?

Also, I was wondering what are you regressing on? i.e. what equation/value are you fitting to?

Italian Stallion
Posts: 35
Joined: Sat Apr 30, 2011 4:42 pm

Re: Reconstructing Box Plus/Minus

Post by Italian Stallion » Sun Jun 23, 2019 11:45 pm

I have no hard data to support this, but I've always felt that all the boxscore models overrate players that score a lot of their points off offensive rebounds (typically Cs, but some others). That combination results in both theoretical value (the offensive rebound) and actual value (the player scored) to be given to the player on what really amounts to a single possession.

When I've looked at boxscore model rankings of players like this and then looked at the evidence from on/off type data, that profile seems to be overrated by boxscore models fairly often.

Imagine, we give the ball to Lebron James in the back court, he breaks a press, drives to the hoop, and uses his strength and athleticism to score.

Imagine, player X misses a shot, Stevie Adams or Enes Kanter is standing under the basket, they get the offensive rebound and put it back in.

In both cases the team has 2 points, but most boxscore models will say that what Kanter and Adams did was worth more because they get credit for both the rebound and score. I'm calling BS on that. The accounting for possessions has some quirky problems in it and this is one example that I think it not correct.

Mike G
Posts: 4429
Joined: Fri Apr 15, 2011 12:02 am
Location: Asheville, NC

Re: Reconstructing Box Plus/Minus

Post by Mike G » Tue Jun 25, 2019 12:43 am

When your team has missed a shot, you no longer have possession. If you retrieve the O-Reb, you have regained possession. If you then score, you should get credit for both.

It's almost like a steal and score, when you neutralize the opponent's possession and also get the points. This is clearly more accomplished than just scoring when your team already has the ball.

Whether you out-rebound your opponent by leaping higher, or you wait for him to come down with the ball and strip him, it's the same result. One is an O-Reb, the other is a steal; one is considered an offensive play, the other defensive. Kind of arbitrary assigning credit to one or the other.

DSMok1
Posts: 905
Joined: Thu Apr 14, 2011 11:18 pm
Location: Maine
Contact:

Re: Reconstructing Box Plus/Minus

Post by DSMok1 » Tue Jun 25, 2019 12:28 pm

It is an interesting discussion!

Ultimately, if offensive rebounders are really not that valuable, a box score regression on a good RAPM basis should show that ORebs are not highly valuable. And to some extent that is the case--Offensive rebounds have a larger P-Value than just about every other typical box score stat. Still, they are a net positive. In my recent regressions, Offensive rebounds typically settle in at worth about 0.18 points TO THE PLAYER. Not a lot, but it is something.

For a team level, we can estimate it directly.

After a missed shot, there is about a 30% chance of an offensive rebound. Possession of the ball is worth about 1 point.

So the expected value for the offensive team is +0.3 points when the ball bounces off the rim. The offensive rebounder gets the ball, and the expected value for the team is now +1.0 points (in reality, it's probably higher than 1.0 due to likelihood of putbacks). The transition is +0.7 points. However, that should be credited half to the offense and half to the defense (in an APM context). Thus the actual value to credit to the offensive team is +0.35 or so.

Since the player-level regression produced a value of +0.18, we should just take the remaining value and credit it to the other 5 players on the court. So the offensive rebounder gets +0.18 and the other 4 players get +0.04.
Developer of Box Plus/Minus
APBRmetrics Forum Administrator
GodismyJudgeOK.com/DStats/
Twitter.com/DSMok1

Mike G
Posts: 4429
Joined: Fri Apr 15, 2011 12:02 am
Location: Asheville, NC

Re: Reconstructing Box Plus/Minus

Post by Mike G » Tue Jun 25, 2019 3:08 pm

Supposing 5 players all go the same 40 or 48 minutes and get varying number of offensive rebounds; and we credit them with the suggested values of .18 when they get the OReb, .04 when a teammate gets it -- they add up like this:

Code: Select all

pos:     C     PF     SF     SG     PG    Tot
OReb:    4      3      2      1      0     10
@0.18  0.72   0.54   0.36   0.18   0.00   1.8
@0.04  0.24   0.28   0.32   0.36   0.40   1.6
tot:   0.96   0.82   0.68   0.54   0.40   3.4
Or you could double these to get the Off+Def contributions?

DSMok1
Posts: 905
Joined: Thu Apr 14, 2011 11:18 pm
Location: Maine
Contact:

Re: Reconstructing Box Plus/Minus

Post by DSMok1 » Tue Jun 25, 2019 3:28 pm

Mike G wrote:
Tue Jun 25, 2019 3:08 pm
Supposing 5 players all go the same 40 or 48 minutes and get varying number of offensive rebounds; and we credit them with the suggested values of .18 when they get the OReb, .04 when a teammate gets it -- they add up like this:

Code: Select all

pos:     C     PF     SF     SG     PG    Tot
OReb:    4      3      2      1      0     10
@0.18  0.72   0.54   0.36   0.18   0.00   1.8
@0.04  0.24   0.28   0.32   0.36   0.40   1.6
tot:   0.96   0.82   0.68   0.54   0.40   3.4
Or you could double these to get the Off+Def contributions?
That looks right to me! The opponents get the other half of the credit.

This is actually how the Box Plus/Minus team adjustment would work when applied at the possession level.
Developer of Box Plus/Minus
APBRmetrics Forum Administrator
GodismyJudgeOK.com/DStats/
Twitter.com/DSMok1

Italian Stallion
Posts: 35
Joined: Sat Apr 30, 2011 4:42 pm

Re: Reconstructing Box Plus/Minus

Post by Italian Stallion » Sat Jun 29, 2019 6:18 pm

Mike G wrote:
Tue Jun 25, 2019 12:43 am
When your team has missed a shot, you no longer have possession. If you retrieve the O-Reb, you have regained possession. If you then score, you should get credit for both.

It's almost like a steal and score, when you neutralize the opponent's possession and also get the points. This is clearly more accomplished than just scoring when your team already has the ball.

Whether you out-rebound your opponent by leaping higher, or you wait for him to come down with the ball and strip him, it's the same result. One is an O-Reb, the other is a steal; one is considered an offensive play, the other defensive. Kind of arbitrary assigning credit to one or the other.
No doubt an offensive rebound is a positive.

No doubt a score is a positive.

I don't think when they occur in combination they are worth more than if Lebron James goes coast to coast.

While I understand the accounting for possessions, IMO there is a flaw in the thinking somewhere that overrates the OREB/Score combination in boxscore models.

Technically, every time you pass the ball, dribble it off your foot etc.. no one actually has possession of it. Keeping it doesn't count as a positive for anyone.

It feels more like it should be (very rough approximation of what I m saying)

If you score you get 2.

If you get an OREB and pass it out you get 1.

If you get an OREB and score yourself you get 2.

Mike G
Posts: 4429
Joined: Fri Apr 15, 2011 12:02 am
Location: Asheville, NC

Re: Reconstructing Box Plus/Minus

Post by Mike G » Sun Jun 30, 2019 12:23 am

I don't think when they occur in combination they are worth more than if Lebron James goes coast to coast.
Does "coast to coast" imply he got the defensive rebound and subsequently scored an unassisted FG?
If so, then it's probably equal to getting an offensive rebound and scoring.

But if it's " we give the ball to Lebron James in the back court" and he scores, that is not equal to a rebound + a score.
Unless it's some LeBron bonus, because he's just cooler.

Nate
Posts: 131
Joined: Tue Feb 24, 2015 2:35 pm

Re: Reconstructing Box Plus/Minus

Post by Nate » Tue Jul 02, 2019 2:46 pm

Italian Stallion wrote:
Sun Jun 23, 2019 11:45 pm
....
Imagine, player X misses a shot, Stevie Adams or Enes Kanter is standing under the basket, they get the offensive rebound and put it back in.

In both cases the team has 2 points, but most boxscore models will say that what Kanter and Adams did was worth more because they get credit for both the rebound and score. I'm calling BS on that. The accounting for possessions has some quirky problems in it and this is one example that I think it not correct.
People typically penalize "player X" for missing, and then offset that penalty with more credit for Adams/Kanter so that the sum of "player X misses" and "Kanter rebounds and scores" is roughly the same as "Lebron goes coast to coast." It seems like you're ignoring that "miss" penalty in your reckoning.

Crow
Posts: 6250
Joined: Thu Apr 14, 2011 11:10 pm

Re: Reconstructing Box Plus/Minus

Post by Crow » Mon Jul 29, 2019 4:45 pm

BPM counts assists as valuable, directly and in the versatility term. Should team % fgs assisted while on the court be considered? I'd think so.

Example: Alfonso McKinnie, very low assist rate. But team assists go up with him on court vs. off. Not by much but it went up not down. I am not asserting he helped (that could be case with other examples of ball movers) but is he hurting being low assists himself? As much as BPM thinks? If others share credit of an offensive rebound, what about assists?

Team efg% and ftr on the court vs. off could be considered too. Doing these things may be the way to catch harm from a low assist guy like McKinnie, if you can adjust for what is expected given teammates behavior overall, with and without. RAPM, RPM considers these. If BPM doesn't directly, it does only crudely in the too crude team adjust. BPM isn't going to be as refined an estimator of passing impact as RAPM/RAPM but it could get somewhat better at measuring passing impact with some acknowledgement that passing and assistmaking are roles, unequally shared and that low assistmaking because of role may not be harmful or particularly harmful as old BPM thinks in extreme cases.

If the AST * REB is removed or substantially changed in importance, there is the opportunity / necessity to redo how assist behavior / value is credited and debited.

DSMok1
Posts: 905
Joined: Thu Apr 14, 2011 11:18 pm
Location: Maine
Contact:

Re: Reconstructing Box Plus/Minus

Post by DSMok1 » Tue Jul 30, 2019 2:31 pm

Crow wrote:
Mon Jul 29, 2019 4:45 pm
BPM counts assists as valuable, directly and in the versatility term. Should team % fgs assisted while on the court be considered? I'd think so.

Example: Alfonso McKinnie, very low assist rate. But team assists go up with him on court vs. off. Not by much but it went up not down. I am not asserting he helped (that could be case with other examples of ball movers) but is he hurting being low assists himself? As much as BPM thinks? If others share credit of an offensive rebound, what about assists?

Team efg% and ftr on the court vs. off could be considered too. Doing these things may be the way to catch harm from a low assist guy like McKinnie, if you can adjust for what is expected given teammates behavior overall, with and without. RAPM, RPM considers these. If BPM doesn't directly, it does only crudely in the too crude team adjust. BPM isn't going to be as refined an estimator of passing impact as RAPM/RAPM but it could get somewhat better at measuring passing impact with some acknowledgement that passing and assistmaking are roles, unequally shared and that low assistmaking because of role may not be harmful or particularly harmful as old BPM thinks in extreme cases.

If the AST * REB is removed or substantially changed in importance, there is the opportunity / necessity to redo how assist behavior / value is credited and debited.
Since I have decided that BPM will not include PbP/Lineup data (for ease of use in older or non-NBA settings), I can't include on-court/off-court data.

The assist*reb term will definitely not be included in the new version of BPM, at least in its present form. It is very bad in dealing with outliers. Some sort of modification will be required.
Developer of Box Plus/Minus
APBRmetrics Forum Administrator
GodismyJudgeOK.com/DStats/
Twitter.com/DSMok1

Mike G
Posts: 4429
Joined: Fri Apr 15, 2011 12:02 am
Location: Asheville, NC

Re: Reconstructing Box Plus/Minus

Post by Mike G » Tue Jul 30, 2019 9:46 pm

If you put an exponent less than 1 on that Ast*Reb, you would reduce the outlier exaggeration.

nbacouchside
Posts: 114
Joined: Sun Jul 14, 2013 4:58 am
Contact:

Re: Reconstructing Box Plus/Minus

Post by nbacouchside » Thu Aug 01, 2019 3:45 pm

Hi Daniel,

Know you've been saying that you don't want to use biometric or pbp stats in the new BPM, but given ability to estimate some things using box-score data with relative accuracy (Ben Taylor's passer rating, offensive load, box creation, etc.) it might be worth it to try to incorporate those sorts of estimates.

Another example of estimating certain categories of talent and even height can be found here: https://nicidob.github.io/automatic_bbgm/ (obviously this is for a simulation but the methodology for producing estimates of things like height seems pretty intuitive).

Think it also might be possible to use box-score information to get a fairly accurate estimate of Games Started % that could be incorporated with MPG (although a big component of GS% would probably be MPG itself).

Italian Stallion
Posts: 35
Joined: Sat Apr 30, 2011 4:42 pm

Re: Reconstructing Box Plus/Minus

Post by Italian Stallion » Wed Aug 14, 2019 3:49 pm

Nate wrote:
Tue Jul 02, 2019 2:46 pm
Italian Stallion wrote:
Sun Jun 23, 2019 11:45 pm
....
Imagine, player X misses a shot, Stevie Adams or Enes Kanter is standing under the basket, they get the offensive rebound and put it back in.

In both cases the team has 2 points, but most boxscore models will say that what Kanter and Adams did was worth more because they get credit for both the rebound and score. I'm calling BS on that. The accounting for possessions has some quirky problems in it and this is one example that I think it not correct.
People typically penalize "player X" for missing, and then offset that penalty with more credit for Adams/Kanter so that the sum of "player X misses" and "Kanter rebounds and scores" is roughly the same as "Lebron goes coast to coast." It seems like you're ignoring that "miss" penalty in your reckoning.
I understand that "player X" missing is some kind of negative.

I am saying people are so worried about the accounting looking right, they are getting the reality of the values wrong.

To be clear, I don't have some kind of fool proof way to account for these things and get the values right. I simply think a LOT of things of value (both positive and negative) are going on in every possession, many of which are not being accounted for, and some that are being given too much value to make the accounting look neat. IMO, the OREB/Putback Score combination is overvalued.

Post Reply