APBRmetrics

The discussion of the analysis of basketball through objective evidence, especially basketball statistics.
It is currently Sat Jun 25, 2016 9:12 pm

All times are UTC




Post new topic Reply to topic  [ 40 posts ]  Go to page 1, 2, 3  Next
Author Message
PostPosted: Mon Oct 29, 2012 10:12 am 
Offline

Joined: Fri Apr 15, 2011 8:28 am
Posts: 780
I've thought about a method to, sort of, compute RAPM for players of the 90s, using

minutes played by each player,
game pace,
both teams ORtg (don't need DRtg because it's equivalent to the opponent's ORtg)

.. for every game played.


My idea was to create fake matchupfiles in the following way:
(pace) times do:
create a line in the matchupfile with 1 possession for each team, teamApoints = teamA_ORtg/100, teamBPoints = teamB_ORtg/100

What's missing? The players! We don't know for sure who was on the court, so we'll simulate
Each player has a chance of MP/48 to appear in each line. So a player with 24 MP will, on average, appear in every second line, and will have played half of the possessions


The thought behind it is this: if a good player plays lots of minutes, it should have a positive influence on the outcome of the game, and vice versa. This method will, most likely, rate those players highly that played many minutes in blowout wins and were missing in blowout losses
This is obviously far from perfect. If someone has an idea to refine the method, I'm all ears

I'll probably post some results in a couple of hours. We can compare those to the actual RAPM numbers we have for the years>2000

_________________
http://stats-for-the-nba.appspot.com/


Top
 Profile  
 
PostPosted: Mon Oct 29, 2012 11:18 am 
Offline

Joined: Fri Apr 15, 2011 12:02 am
Posts: 3306
Location: Asheville, NC
J.E. wrote:
minutes played by each player,
game pace,
both teams ORtg (don't need DRtg because it's equivalent to the opponent's ORtg)

.. for every game played.

This is a very nice idea.
Would it be even better to use both teams' SRS for the season, minus their (ORtg - DRtg) for the game, rather than just ORtg for the game?

In the past we've argued about a player's worth by noting facts like, "They were 40-30 with him in the lineup, and 5-7 without him".
The rejoinder might be, "Yes, but they were 5-7 against better than average teams".
Then we'd introduce point differentials to further refine the arguments.

If you compare [expected pt-diff via SRS] to [actual pt-diff], giving players credit according to their minutes, you've surely got something.
Even blowouts would reveal something. You might scale them back somewhat.


Top
 Profile  
 
PostPosted: Mon Oct 29, 2012 12:08 pm 
Offline

Joined: Mon Apr 18, 2011 10:09 am
Posts: 458
Have you tested that concept with the current data and the compared it to the real matchup files? Reverse engineering should give you the best approach.

Well, also, you can use the starting lineups as point of reference.

_________________
http://bbmetrics.wordpress.com/


Top
 Profile  
 
PostPosted: Mon Oct 29, 2012 2:25 pm 
Offline

Joined: Fri Apr 15, 2011 8:28 am
Posts: 780
mystic wrote:
Have you tested that concept with the current data and the compared it to the real matchup files?
what exactly would I look for when I compare them?
Quote:
Well, also, you can use the starting lineups as point of reference.
That's a good idea; I could force ~1/4 of all possessions to include starters only (assuming the starters play 6 minutes against each other at the start of the 1st and the 3rd quarters)

Mike G wrote:
Would it be even better to use both teams' SRS for the season, minus their (ORtg - DRtg) for the game, rather than just ORtg for the game?
I'm guessing (hoping) that RAPM will do the strength of schedule-adjustment for us.
Say, if Jordan in 97+98 scores a good DRtg (opp ORtg) against the Jazz in many games, and the Jazz have a generally good ORtg in most other games, Jordan's rating should get an extra boost

_________________
http://stats-for-the-nba.appspot.com/


Top
 Profile  
 
PostPosted: Mon Oct 29, 2012 3:32 pm 
Offline

Joined: Fri Apr 15, 2011 8:28 am
Posts: 780
R^2 for real RAPM and fake RAPM is ~0.25. Fake RAPM has everyone even closer together than RAPM does, so it's really hard to get a good rating when being on a bad team etc.

The top ~10 for 2011 look like this:

Dirk Nowitzki (+2.3)
Semih Erden(?)
Chris Bosh
Luol Deng
Omer Asik
LeBron James
Tyson Chandler
Nick Collison
Kenyon Martin
Joel Anthony
Richard Jefferson
Carlos Arroyo
Manu Ginobili (+1.6)

Not horrible, but not great either. Better than nothing I guess. I'll see if _forcing starters to play against each other more_ changes something

_________________
http://stats-for-the-nba.appspot.com/


Top
 Profile  
 
PostPosted: Mon Oct 29, 2012 4:02 pm 
Offline

Joined: Mon Apr 18, 2011 10:09 am
Posts: 458
Well, looks like you've done already, what I thought about. Create the fake matchup file, then run RAPM algorithm, then compare fake RAPM with real RAPM.

The other point would be lineups which can be compared. How many minutes are getting certain lineups in your fake matchup file and compare that to the real minutes the lineups got. Maybe that gives you something to work with in terms of adjustments.

Also, the expected ORtg by a team would be based on the own ORtg and the opponents DRtg. If I have a team with 110 ORtg playing a 100 DRtg team, I expect 105 as the result, while against a 110 DRtg team I would rather expect 110.

Maybe you can also use your boxscore stats model as prior, maybe you would get a closer result to the real RAPM then?

_________________
http://bbmetrics.wordpress.com/


Top
 Profile  
 
PostPosted: Mon Oct 29, 2012 7:51 pm 
Offline

Joined: Sat Apr 16, 2011 7:40 am
Posts: 306
Location: Cambridge, MA
mystic wrote:
Also, the expected ORtg by a team would be based on the own ORtg and the opponents DRtg. If I have a team with 110 ORtg playing a 100 DRtg team, I expect 105 as the result, while against a 110 DRtg team I would rather expect 110.


I'm not sure this is correct?

An ORtg of is vs ~league average (assuming balanced schedule), or what you expect to have vs that league average DRtg. So if that's league average that given year, that works. That all assumes if it's linear, which I'm not sure it is either.

Other words the Thunder last year w/ a 110 ORtg vs the Bobcats w/ a 110 DRtg. We shouldn't expect a 110 ORtg, instead, we'd expect ~115. We'd expect a 110 ORtg vs the Hornets (a league average defense) and something like a 105 ORtg vs a top defense (like the Heat)

_________________
http://pointsperpossession.com/
@PPPBasketball


Top
 Profile  
 
PostPosted: Mon Oct 29, 2012 8:18 pm 
Offline

Joined: Mon Apr 18, 2011 10:09 am
Posts: 458
Bobbofitos wrote:
I'm not sure this is correct?


Well, I'm sure it is incorrect. I wanted to use an example with league average being 105. In that case 110 ORtg vs. 100 DRtg, would be +5 offense vs. +5 defense, which would then equate to 0 or league average as expectation. The other example, as you correctly noted, would be +5 offense vs. -5 defense, which would be +10 or 115 Ortg.

It should be linear, at least in the first order. ;)

_________________
http://bbmetrics.wordpress.com/


Top
 Profile  
 
PostPosted: Tue Nov 06, 2012 10:09 pm 
Offline

Joined: Fri Apr 15, 2011 8:28 am
Posts: 780
I've done some multiyear fake RAPM now, hoping that multiple years of data would help with accuracy. There's one large problem though: Bench players of very good teams get a very high rating. Why? Because they play more minutes when their team's starting 5 (the actual good players) were *so* good that the game was a blowout, which probably led to the starting 5 playing no minutes later in the game. Those minutes were instead played by the second unit. The algorithm associates "winning with large differential" with "extended playing time of bench players" (for good teams), and, not knowing any better, assumes the bench players were the reason for the good differential.

For 1991-1993 Bird, Magic and KJ go #1, #3 and #5, which is nice. Jordan is #8 though. Ranked higher than him are Cliff Levingston (played ~1000 minutes for CHI in 91 and 92), Wayne Cooper (played 1000 minutes total for POR in 91 and 92), Will Perdue (~1000 MP/year for CHI in 91, 92, 93)

Maybe it helps if I use even more years. I'll have to check

_________________
http://stats-for-the-nba.appspot.com/


Top
 Profile  
 
PostPosted: Wed Nov 07, 2012 2:02 pm 
Offline

Joined: Fri Apr 15, 2011 8:28 am
Posts: 780
10 year fake RAPM (91-00) give the following ranking

Jordan (+4.5)
Bird (+4)
Magic (+3.8)
Rodman
Shaq
Pippen
DRobinson
Wayne Cooper
Mourning
Cliff Levingston
Eddie Jones
Doc Rivers
Danny Ainge
Dan Majerle
Sam Perkins
Penny Hardaway (+2.6)

If I ever want to really rate players of that timeframe with some sort of +/- metric I'll probably have to include adjacent seasons (weighing them a little less). So, for rating players in 95 I might use 1.0*95data, 0.8*96data, 0.8*94data, 0.5*97data and 0.5*93data

_________________
http://stats-for-the-nba.appspot.com/


Top
 Profile  
 
PostPosted: Wed Nov 07, 2012 2:59 pm 
Offline

Joined: Sun May 06, 2012 12:18 am
Posts: 20
What data do you have? Just thinking that if you have quarter by quarter totals(Basketball reference seems to have them, just by scanning) you could break down the matchup probabilities into minutes played per quarter by starters, and maybe matchup the general pattern to how teams tend to play starters now(How long does the average starter play 1st quarter vs second quarter, 3rd vs 4th etc). Might (partially) overcome that blowout win victory problem. Say 32min a game starters(now) tend to play 10 in 1st, 6 in second, 8in 3rd and 8 in 4th(Or anything that isn't a constant across quarters) and then reweigh minutes played to account for the distribution of minutes, and then assign starters 10/12*1st quarter differential + 6/12*2nd quarter differential...

Not sure how much that'd help, just trying to think of other ideas for modelling. Fake RAPM seems to be reasonable, Magic, Bird and Jordan at the top isn't laugh inducing. They were pretty good.


Top
 Profile  
 
PostPosted: Wed Nov 07, 2012 3:08 pm 
Offline

Joined: Fri Apr 15, 2011 8:28 am
Posts: 780
dtjmcauliffe wrote:
Just thinking that if you have quarter by quarter totals(Basketball reference seems to have them, just by scanning)
where? for seasons earlier than 2001?

_________________
http://stats-for-the-nba.appspot.com/


Top
 Profile  
 
PostPosted: Wed Nov 07, 2012 3:14 pm 
Offline

Joined: Mon Apr 18, 2011 10:09 am
Posts: 458
J.E. wrote:
dtjmcauliffe wrote:
Just thinking that if you have quarter by quarter totals(Basketball reference seems to have them, just by scanning)
where? for seasons earlier than 2001?


http://www.basketball-reference.com/box ... 10ATL.html

Since 1976/77 they have the score for each quarter.



You could also look how realistic it is to assume that much of the point differential came with the subs, by looking at the available dataset. In that way you may find a pattern to use to predict minute distribution for starters and subs across the quarters dependent on the scoring margin for the start and the end of the respective quarters.

_________________
http://bbmetrics.wordpress.com/


Top
 Profile  
 
PostPosted: Wed Nov 07, 2012 3:52 pm 
Offline

Joined: Fri Apr 15, 2011 8:28 am
Posts: 780
mystic wrote:
J.E. wrote:
dtjmcauliffe wrote:
Just thinking that if you have quarter by quarter totals(Basketball reference seems to have them, just by scanning)
where? for seasons earlier than 2001?


http://www.basketball-reference.com/box ... 10ATL.html

Since 1976/77 they have the score for each quarter.
Ah. I was kinda hoping he meant minute totals for each quarter. Quarter by quarter point totals are obviously useful, too. I guess now I'll have to do some research on starter/bench minute distribution according to quarter

_________________
http://stats-for-the-nba.appspot.com/


Top
 Profile  
 
PostPosted: Wed Nov 07, 2012 9:16 pm 
Offline

Joined: Fri Apr 15, 2011 12:02 am
Posts: 3306
Location: Asheville, NC
Quote:
Bench players of very good teams get a very high rating. Why? Because they play more minutes when their team's starting 5 (the actual good players) were *so* good that the game was a blowout,...

In a blowout, especially one in which the 2nd string gets lots of minutes, it's often a lopsided score after 3 quarters.

Maybe in such games, you should judge the subs by how the 4th quarter went, more than by the whole game pt-diff.


Top
 Profile  
 
Display posts from previous:  Sort by  
Post new topic Reply to topic  [ 40 posts ]  Go to page 1, 2, 3  Next

All times are UTC


Who is online

Users browsing this forum: Yahoo [Bot] and 8 guests


You cannot post new topics in this forum
You cannot reply to topics in this forum
You cannot edit your posts in this forum
You cannot delete your posts in this forum

Search for:
Jump to:  
Powered by phpBB® Forum Software © phpBB Group