The debut and popularization of BPM

[Mike G]

Neil, what stats do you need?
Minutes, team, eWins per minutes? For players switching teams, per team and/or combined?

[mystic]

Has everyone seen this visualization? http://public.tableausoftware.com/share ... _count=yes

Thanks, didn't see that before.

I tried to highlight the weak areas in the writeup. http://www.basketball-reference.com/about/bpm.html

I didn't read that word for word, but I think you did a fine job pointing out weaknesses.

[Neil Paine]

Does anyone think the players at left would be competitive with those on the right?

That's somewhat disingenuous, Mike -- the team on the right would have a higher BPM as well, +0.2 to -0.7. Granted, that gap works out to about 3 wins per 82 games, but it's not like the team on the left is considered "better" than the one on the right, even after cherry-picking an extreme example.

Code: Select all

+-----------------------+------+------+--+-------------------+------+------+
|       BPM Stars       |  MP  | BPM  |  |    Mike's Team    |  MP  | BPM  |
+-----------------------+------+------+--+-------------------+------+------+
| Kyle Singler          | 2337 | -0.2 |  | Monta Ellis       | 3023 | 0.2  |
| Wesley Johnson        | 2240 | 0.1  |  | Greg Monroe       | 2690 | 0.2  |
| Martell Webster       | 2157 | 0.0  |  | Joe Johnson       | 2575 | -0.1 |
| Iman Shumpert         | 1962 | 0.1  |  | LaMarcus Aldridge | 2498 | 0.1  |
| Maurice Harkless      | 1950 | -0.2 |  | Reggie Jackson    | 2277 | -0.2 |
| Mike Miller           | 1707 | 0.2  |  | Jared Sullinger   | 2041 | -0.1 |
| Dante Cunningham      | 1635 | -0.2 |  | Nikola Pekovic    | 1663 | 0.2  |
| Caron Butler          | 1419 | -2.1 |  | Nene Hilario      | 1560 | 0.0  |
| Francisco Garcia      | 1083 | -0.1 |  | Kyle O'Quinn      | 1188 | 0.2  |
| Chris Douglas-Roberts | 1016 | 0.1  |  | Rajon Rondo       |  998 | 0.1  |
| Dorell Wright         |  984 | -0.2 |  |                   |      |      |
| Rashard Lewis         |  971 | 0.2  |  |                   |      |      |
| Andrei Kirilenko      |  857 | 0.0  |  |                   |      |      |
| Jeff Withey           |  684 | 0.0  |  |                   |      |      |
| Andre Roberson        |  399 | 0.0  |  |                   |      |      |
+-----------------------+------+------+--+-------------------+------+------+
| TOTAL                 |      | -0.7 |  | TOTAL             |      | 0.2  |
+-----------------------+------+------+--+-------------------+------+------+
| WINS/82               |      | 39.4 |  | WINS/82           |      | 41.6 |
+-----------------------+------+------+--+-------------------+------+------+

Neil, what stats do you need?
Minutes, team, eWins per minutes? For players switching teams, per team and/or combined?

Ideally I'd have player name, year, minutes and eWins… Doesn't matter if multi-team guys are split by team or combined in a row. And if you could match up to Basketball-Reference ID, that would be even better.

[Mike G]

Well, you've got Butler at -2.1, while I had sorted him by his OKC minutes at 0.2 BPM. So they were all supposed to be near-0 players. Even if one side avg +0.1 and the other side were all -0.1, there would be just a 1.0 ppg difference.

Is this cherry picking? We can take any BPM stratum with ~40 players and sort them by eWins and get a similar disparity:

Code: Select all

BPM 0.6 to 0.9           eWins
Klay Thompson		David Lee
Cory Joseph		Zach Randolph
Mike Dunleavy		DeMar DeRozan
Randy Foye		Kenneth Faried
Courtney Lee		Tyreke Evans
Jodie Meeks		Marcin Gortat
Jeff Ayres		D.J. Augustin
Nick Collison		Jrue Holiday
Jae Crowder		Markieff Morris
Thabo Sefolosha		
Greg Stiemsma		
Shane Battier		

Both teams total about 37 WAR (BPM); the eWins team totals 59 eW to the others' 28; WS are 54 to 43; PER averages are 18.2 and 12.2 (not minutes weighted).
By eWins, only Thompson is above-avg in the left column. The rest are .15 to .86 of avg.

Hand entering all b-r.com player ID is a big job. Do you have an easier way to do it?

[Neil Paine]

The "better" team is still better by BPM, but that's beside the point… If this trend is legitimate (i.e., BPM is systematically overvaluing players), then eWins will be a better predictor of actual NBA teams' performance. If not, then just chalk it up to the cherry-picking exercise, and know that the criticism doesn't hold up empirically.

Hand entering all b-r.com player ID is a big job. Do you have an easier way to do it?

You can dump the names into http://www.basketball-reference.com/friv/linkify.cgi and replace the names with URLs (which can then be stripped to the IDs). Or I can just do it myself.

[Crow]

DSMok1, wins produced and per don't have to perform well out of sample on their own to nudge a metric blend to a more optimal level. But there might be other candidates that help on same stats. Might be worth looking at adding extra helping of one or more of the rapm factor level components to the blend to balance for what the box score metrics may do poorly, don't do at all or just enough. I.e., defensive rebounding impact on teammates and adj. PPS on offense instead of WP and PER weights and defensive adj. PPS to backfill for weaknesses. If Neil sets up all the data, can search for his best performing blend(s).

[DSMok1]

DSMok1, wins produced and per don't have to perform well out of sample on their own to nudge a metric blend to a more optimal level. But there might be other candidates that help on same stats. Might be worth looking at adding extra helping of one or more of the rapm factor level components to the blend to balance for what the box score metrics may do poorly, don't do at all or just enough. I.e., defensive rebounding impact on teammates and adj. PPS on offense instead of WP and PER weights and defensive adj. PPS to backfill for weaknesses. If Neil sets up all the data, can search for his best performing blend(s).

Correct, Crow. Just because a metric doesn't perform as well does not mean it won't add information. We would have to be careful of overspecifying if we use blends (need cross validation).

In fact, getting back to Mike G's discussion of taking average players with one metric and splitting them into two teams with another metric:

As long as the two metrics are not perfectly correlated and they both have some level of validity, the effect will always be the same. If I take average players with one metric and split into two teams via another valid metric, the team better via that second metric will always be better. Period. If I take average players with eWins and split into two teams via BPM, the better team via BPM will be better than the worse team via BPM.

Why? It comes back to the Bayesian discussion above, and what Crow said here. We are taking one set of "information" about a players' quality and adding more information. As long as that information has some validity, it will always help refine our estimate of who's better!

[Mike G]

If I take average players with eWins and split into two teams via BPM, the better team via BPM will be better than the worse team via BPM.

That's a beautiful thought. Is one of these teams clearly better than the other?

Code: Select all

Josh Smith          Klay Thompson
Thaddeus Young     Wesley Matthews
Tristan Thompson     Robin Lopez
Jameer Nelson        Joe Johnson
Dion Waiters        Jose Calderon
Miles Plumlee      Marco Belinelli
Nick Young          Chris Andersen
Mike Scott           Tony Allen
Omer Asik            C.J. Miles
Ed Davis            Nate Robinson
Ramon Sessions      

Both teams totaled about 19,900 minutes last year. All were rated from .95 to 1.05 of avg in eWins.
One team totaled 42 eWins, the other 41. Avg PER concurs, 15.9 to 15.2
That slightly better team had 54.5 Win Shares, to 30.4 for the other. And a 45-8 disparity in WAR (via BPM)

[DSMok1]

Yes, that's my point. In this case, the second team is way better than the first team--similar to how the eWins split the BPM average players. Not all of the second team was above average in BPM, though--Nate Robinson and Joe Johnson weren't.

Yes, I'd take Tony Allen over Nick Young.

[mystic]

Mike, your way of trying to prove your idea with those hand-picked examples is a bad way of arguing. You are running into a confirmation bias here, and that is never useful.

In average BPM performed extremly well in retrodiction tests. Let us how your eWin metric performs and then we can judge. You feel your metric is somehow better at evaluating players, but so far that is only your feeling and no matter how many cherry-picked examples you can find, that will not change.

Also, something I think needs to put out there more often, the boxscore values players produce are not just based on an intrinsic player skill, but effected by the circumstances. Having a list of player values generated in those circumstances and then put those players into different circumstances will always look like those rather "role players" are worse than the metric tells us. But in reality, if we evaluate the "star-players" in the same way, we see that not only roleplayers have trouble to generate the same numbers in different roles, but even star players can get marginalized by a lesser role. Take the 2011 Heat as an example or how the last couple of years had an influence on most people's opinion on Chris Bosh (and your eWins as well), now he looks completely different again for most people despite the fact that he is basically the same player as he was before except of him playing a different role. The point is: People are quick at recognizing that role players likely have trouble adopting to a bigger role, but they ignore that star players often also have trouble to adjust to lesser roles.

[Mike G]

Here are players in the range of 1.15 to 1.25 eW/484. One team has all positive BPM (up to +4.2); the other runs from 0.1 down to minus-1.9

Code: Select all

Brandon Knight      Monta Ellis
Taj Gibson         Nicolas Batum
Reggie Jackson    Chandler Parsons
Kevin Martin        Ricky Rubio
J.J. Hickson      Markieff Morris
John Henson       Anderson Varejao
Tobias Harris      Mason Plumlee
Kosta Koufos        J.J. Redick
Rajon Rondo        James Johnson
Ryan Anderson      Gorgui Dieng
Jeff Adrien      
Jon Leuer      

Remembering that we are thinking of last seasons' performances. Is there an obvious better team?
This time, both squads totaled 49 eWins last year. Both were just under 17 avg PER.
Win Shares ran 52 to 41; WAR total 53 to 17.

[Mike G]

There's nothing hand or cherry picked about these 'teams'. In fact, it's the exact opposite. As DSMok1 suggested, I took 20 players from the 'average' zone of e484, sorted them by BPM, and put all the better BPM's on one team.
I repeated the experiment with another zone, of above avg eWins guys.

In both cases, PER agrees that the teams are ~equal. Win Shares seemed to be closer to eWins (than to WAR) in the 2nd example.

I don't even know what 'my idea' would be that I'm trying to 'prove'. These are open experiments: You may conclude what you will.
However, I do feel that the teams created from nearly-equal BPM -- sorted by e484 -- are less equal than those from nearly-equal e484 players.

In these last 2 examples, everyone on one side is better than anyone on the other side, according to BPM. Yeah, Tony Allen probably is worth more to the Grizzlies; but Nick Young might be worth more to the Lakers.
In fact, total salaries of these teams are about equal.

[DSMok1]

In both of these split sample cases, WS/48, BPM, xRAPM, and RAPM all agree that one team is better than the other. You can do the same thing with probably any stat.

[colts18]

dsmok1, have you tried experimenting with adjusting players BPM scores by teams 4 factors rankings? Lets say a defense is top 10 because its strong in forcing turnovers (LeBron's Heat) and lower eFG% at the expense of rebounding, shouldn't the perimeter guys get more credit for that result? If the defense is strong because of defensive rebounding and low eFG% (Spurs), then the big men should get more credit from BPM.

[DSMok1]

dsmok1, have you tried experimenting with adjusting players BPM scores by teams 4 factors rankings? Lets say a defense is top 10 because its strong in forcing turnovers (LeBron's Heat) and lower eFG% at the expense of rebounding, shouldn't the perimeter guys get more credit for that result? If the defense is strong because of defensive rebounding and low eFG% (Spurs), then the big men should get more credit from BPM.

I have not tried that. That's a bit outside the scope of what I want the stat to be, which is as simple and streamlined as possible.