 |
APBRmetrics
The statistical revolution will not be televised.
|
| View previous topic :: View next topic |
| Author |
Message |
Dan Rosenbaum
Joined: 03 Jan 2005
Posts: 541
Location: Greensboro, North Carolina
|
 Posted: Wed Jan 19, 2005 2:52 pm Post subject: Do Roland Ratings overcorrect for off-court +/-? |
 |
|
My understanding is that this is the definition of the Roland Rating.
Roland Rating = (on-court +/-) minus (off-court +/-)
These are all per-48 minute measures.
Wouldn't it be better if the definition were the following?
Roland Rating = (on-court +/-) minus (off-court +/-) plus (20% of average team +/-)
Alternatively, one could just normalize the Roland Rating so that the average minutes-weighted Roland Rating for a given team was equal to the team's average plus/minus.
The Roland Rating right now only measures a player's value to his team relative to some sort of an average for that team. Thus, a team of clones who were all +4 per 48 minutes would have a Roland Rating of 0 despite the fact that the team won games by an average of 20 points per 48 minutes.
Adding back in the team plus/minus, I think, would make the inter-team comparisons more meaningful. |
|
| Back to top |
|
 |
stuart mckibbin
Joined: 13 Jan 2005
Posts: 17
|
| Posted: Wed Jan 19, 2005 9:34 pm Post subject: |
|
|
I do not know why Roland gives the offcourt number and the oncourt number equal weight when he adds them algebraically. For example, about a week ago, Bryant was on the court 1382 minutes and offcourt 161 minutes. Oncourt +/- was +2.5 and the offcourt +/- was -13.7. The vast bulk of his Roland rating was based on just 161 minutes of offcourt time, making it much more subject to random chance. One simplistic way to correct for it would be to prorate the oncourt and offcourt numbers to total minutes using this formula in Bryant's case
Roland Rating = ((oncourt+/-)*1382/1543) minus ((offcourt+/-)*161/1543)
The Roland Rating right now only measures a player's value to his team relative to some sort of an average for that team.
In Bryant's case I think Roland's rating may only capture the difference between a him and his scrub replacement, rather than a team average.
Dan, I like your idea of adding back 20% of the team's +/-. I would also suggest to make interteam comparison more meaningful the offensive and defensive 100 possession numbers should be used instead of per 48 minute numbers. |
|
| Back to top |
|
|
Dan Rosenbaum
Joined: 03 Jan 2005
Posts: 541
Location: Greensboro, North Carolina
|
| Posted: Wed Jan 19, 2005 11:49 pm Post subject: |
|
|
I think Roland has to give the on-court and off-court numbers the same weight. Theoretically, I am not sure how we would interpret the rating if he did not do so. But you are right that players that play almost all of the time are going to have very imprecisely measured Roland Ratings.
I do very much like the idea of making the measure a per 100 possessions measure rather than per 48 minutes. In general, anytime we are able to remove the effect of pace, I think it is a very good idea. Very good thinking on this. |
|
| Back to top |
|
|
Roland_Beech
Joined: 14 Jan 2005
Posts: 43
|
| Posted: Thu Jan 20, 2005 1:00 am Post subject: |
|
|
I deny I ever created these ratings...
My name has been falsely attached to them...
I have been framed I tell you!
Actually I am pretty much at the stage of wanting to de-emphasize the on court/off court numbers on the site, however they are the most popular feature with your casual NBA follower, or even a fairly serious NBA follower who is not the stats buff that we all here are. So alas, they continue on in their imperfect form.
Part of the reason I've left them alone so far is that they do have a fairly 'easy to understand' convention: on court is the team's net points per 48 minutes with the guy, off is the net without the guy, and the difference is the -eek- Roland Rating.
I heartily agree with Stuart's comment that they are of little reliability for players with either large or small playing time...they are better suited for those players who have been on court for 50% of a teams total activity... if only coaches would get with the program.
I don't believe however that the weighted suggestion really solves the problems -- then you are rendering the on/off rating to be practically a hybrid of a player's raw +/- and is too dependent on the overall quality of the player's team. I.E. there's no way a guy like Marbury playing on a <.500 team like the Knicks can keep pace with any of the Spurs regulars...
A similar issue is what to do with the games a player missed entirely, or the games played before a player arrived on the team in a trade.
As far as going to a possession based number, we already post those on the individual player on/off pages (for instance, see Brian Cardinal at http://www.82games.com/04MEM11D.HTM and yes I believe they are superior. Again though, it's harder to explain the notion of possession ratings to a first time visitor to the site looking at his first 82games page...
As for the original suggestion of adding 20% of the team's overall +/- per 48 minutes, in most cases this will have a very minor effect -- if a team is +2 net per 48, then you're talking about a 0.4 adjustment on ratings that by the end of the season are typically from +10 to -10. At the same time, adding +2.2 to each San Antonio player's ranking seems a little extreme (and yet wouldn't help Malik Rose, currently sporting a -15.3 on/off rating despite a respectable +/- of +14 in his minutes)
So the bottom line is they are what are they are -- less interesting than DanVal (for which we will hopefully have an update using the current 04-05 season soon) but not a bad first impression perhaps of which players seem to be making a difference for a team.
I am certainly open to ideas for improving them so long as the adjustments justify any additional explanations needed for how the ratings come to be. The most likely next step would be to publish the more simplistic (than DanVal) adjusted versions for the other players on the court...but that falls behind about thirty research projects in my priority list.[/url] |
|
| Back to top |
|
|
Kevin Pelton
Site Admin
Joined: 30 Dec 2004
Posts: 979
Location: Seattle
|
| Posted: Thu Jan 20, 2005 2:41 am Post subject: Re: Do Roland Ratings overcorrect for off-court +/-? |
|
|
| Dan Rosenbaum wrote: | | Thus, a team of clones who were all +4 per 48 minutes would have a Roland Rating of 0 despite the fact that the team won games by an average of 20 points per 48 minutes. |
If I'm reading your example right, they'd be +4 per 48 minutes as a team, right? So the difference isn't that dramatic.
| Quote: | | In Bryant's case I think Roland's rating may only capture the difference between a him and his scrub replacement, rather than a team average. |
Well, you're not really able to play "team average" when Bryant goes down with an injury; better players than said scrub replacement will get time, but still assumedly worse ones than average because of team construction. You don't get really good backups to your best players. (Which means their value is limited if they're injury-prone, but that's a bigger team-value issue.)
The fact that teams get killed with guys like Bryant and Garnett off the court, while it stands a higher chance of being affected by randmoness and a couple of unimportant stretches, still has intellectual value. One of the neat things about the off-court data to me is seeing just how important it really is when announcers talk about playing even without a star player on the court. Teams that can do that derive a big advantage.
I think adding in team quality as Dan suggests has value in giving proper credit to good teams, but I think I'd prefer it as a complement to the current Roland Ratings rather than a replacement to them. |
|
| Back to top |
|
|
kjb
Joined: 03 Jan 2005
Posts: 865
Location: Washington, DC
|
| Posted: Thu Jan 20, 2005 9:03 am Post subject: |
|
|
To me, the Roland ratings are like a signpost. If I see what I think is an aberrant +/- or something that doesn't fit with what I'm thinking from watching games, I know it's time for investigation. I think they're a reasonable, straightforward and easily explained first cut at player value. There might be some adjustments that could make them better, but I worry that as the explanation for the rankings increases in complexity, the usefulness to fans will decrease.
For the basketball sabermetrician, I don't think it's much of an issue because Roland is generous with the raw +/- data. That enables us to make whatever adjustments we want in our analysis. |
|
| Back to top |
|
|
Mike G
Joined: 14 Jan 2005
Posts: 3608
Location: Hendersonville, NC
|
| Posted: Thu Jan 20, 2005 9:23 am Post subject: |
|
|
| Quote: | | [Kevin P:] I think adding in team quality as Dan suggests has value in giving proper credit to good teams, but I think I'd prefer it as a complement to the current Roland Ratings rather than a replacement to them. |
If it's no harder to do, why not give team "credit" to players?
If players on a given team only are comparable to their own teammates, how (when) is that better than being comparable to players on other teams, as well as their own?
There's no diminishment of accessibility, and there's less chance of being put off by inexplicable ratings. Good teams have good players. |
|
| Back to top |
|
|
Dan Rosenbaum
Joined: 03 Jan 2005
Posts: 541
Location: Greensboro, North Carolina
|
| Posted: Thu Jan 20, 2005 9:26 am Post subject: Re: Do Roland Ratings overcorrect for off-court +/-? |
|
|
| admin wrote: | | Dan Rosenbaum wrote: | | Thus, a team of clones who were all +4 per 48 minutes would have a Roland Rating of 0 despite the fact that the team won games by an average of 20 points per 48 minutes. |
If I'm reading your example right, they'd be +4 per 48 minutes as a team, right? So the difference isn't that dramatic. |
Roland and I had a discussion about this yesterday (although this does not mean he endorses anything I say here  ). A teamful of +4 clones would win games by 20 points per 48 minutes.
A +4 player who played with a team of 0 rated players would win by 4 points per 48 minutes when he was on the court and his off-court plus/minus would be zero, leading to a +4 Roland Rating. If he played just 24 minutes a night, the team would win games on average by 2 points (4 points per 48 minutes times half the game).
Now suppose we replaced another 0 player with a +4 player who played 24 minutes. Suppose also that he played half his minutes when the other +4 player was on the floor and half when he was not.
The team would win by 8 points per 48 minutes during the 12 minutes the two players shared the floor, 4 points per 48 minutes during the 24 minutes they played separately, and 0 when neither was playing. This would lead to the team winning by 4 points per 48 minutes per game.
We could keep substituting +4 players for 0 players and we would eventually get to the point where the team would win by 20 points and each player would have a Roland Rating of 0.
The point of this example is that if we wanted to use the Roland Ratings to predict how a certain combination of players would do on the floor (versus an average opponent), we would need to add up the Roland Ratings of each of the five players. We would add them up and not average them. (Then we would need to add in an adjustment for the team's average plus/minus.)
There will be cases where this will lead to absurd results. For some teams certain players almost always play together. Take, for example, if a team had five +4 players who always played together for 24 minutes a game and five 0 players who substituted for them. The team would win by an average of 10 points per game and the Roland Rating would be +20 for each of the +4 players. The Roland Ratings, in this case, would be biased because each of the +4 players got to play with much better players than did the 0 players (whose Roland Rating would be -20).
Now this whole discussion has ignored the effect of player combinations for the opponents, which complicates this even further.
The Roland Rating measures the average difference in team performance when a player is on the floor versus when he is not. If the difference in quality of the players that the given player plays with and against is the same both when he is on and off the court, the Roland Rating measures only that given player's impact impact on winning (relative to some team average).
The problem is that this "ceteris paribus" assumption rarely holds, so we are typically measuring some combination of the player's impact and the relative quality of the players he shares the floor with. |
|
| Back to top |
|
|
Kevin Pelton
Site Admin
Joined: 30 Dec 2004
Posts: 979
Location: Seattle
|
| Posted: Thu Jan 20, 2005 11:49 pm Post subject: |
|
|
Is your argument, then, that the players themselves have an inherent value of +4 points per 48 minutes as compared to an average replacement?
That, then, is different than having an "observed" +4 rating.
If you look at the Roland Ratings leaderboard, most of the top players look to be from good teams. (A notable exception is Peja Drobnjak.)
It would be interesting to see a leaderboard with an adjustment for team strength. |
|
| Back to top |
|
|
Dan Rosenbaum
Joined: 03 Jan 2005
Posts: 541
Location: Greensboro, North Carolina
|
| Posted: Fri Jan 21, 2005 6:53 am Post subject: |
|
|
| admin wrote: | Is your argument, then, that the players themselves have an inherent value of +4 points per 48 minutes as compared to an average replacement?
That, then, is different than having an "observed" +4 rating.
If you look at the Roland Ratings leaderboard, most of the top players look to be from good teams. (A notable exception is Peja Drobnjak.)
It would be interesting to see a leaderboard with an adjustment for team strength. |
Yes, I am defining a +4 player such that a team with one +4 player (who played 48 minutes) and the rest of the team 0 players would win games by 4 points. A team full of +4 players would win by 20 points.
And yes, I agree that these +4 players would not always get a Roland Rating of +4. In my examples above, they got ratings of +20, +4, and 0.
As Roland mentioned, adjusting for team plus/minus would not make that much of a difference. |
|
| Back to top |
|
|
|
|
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
You cannot vote in polls in this forum
|
Powered by phpBB © 2001, 2005 phpBB Group
|
FAQ
Search
Memberlist
Usergroups
Register
Profile
Log in to check your private messages
Log in
