APBRmetrics

Yeah, that could be another useful measurement to look at in relation to all the other Adjusted +/= and the raw assist data.

I have to say I don't care too much about assists to spend time on this. I don't like the fact that scorekeepers award them differently and that some players have a way easier time to get assists than others because of who they're playing with.
I'm pretty happy that, with offensive RAPM and impact on points per shot, we have two measures that are not at all influenced by BoxScore assist numbers.

In the meantime I crawled ESPN for PBP files and wrote my own Play By Play-file to matchup-file parser. That means there's some RAPM for '01/02' to '04/'05 coming up. And a big update on coaches, too. Only have 01/02 for now http://stats-for-the-nba.appspot.com/ranking02. Play by Play files from that season are mostly from March through June, they didn't have much for the earlier months. Even those that ESPN has contain many errors which made some quarters of the games unusable. Errors usually come from not listing substitutions.

2002 RAPM likes Doug Christie more than I expected. He was minute leader on the #1 (in SRS) Kings team though.

With less than 1/3 of a season that data probably isn't too useful. It gets alot better for the next 3 years though.

Currently the file lists everybody that played in the last 10 years. Sorry about that, it will be fixed in a couple of days

Wow, that's wonderful news, J.E.! I look forward to the rest of the results.

J.E. wrote:I have to say I don't care too much about assists to spend time on this. I don't like the fact that scorekeepers award them differently and that some players have a way easier time to get assists than others because of who they're playing with.

Jerry, the reason I thought it would be interesting was for exactly this reason. If certain players get assists because they play with certain teammates, I figured it would show up in adjusted assists.

Because scorekeepers are different, you can just use the ratio: Away(Ast/FG)/Home(Ast/FG), as a correction to their assist totals at home.
To adjust for teammates' scoring ability, use the Ast% (% of teammates' FG assisted) formula at b-r.com. Of course, this doesn't differentiate whether your teammates are scoring because you are passing well, or if you are getting assists because they are shooting well.

I updated the multiyear ratings, which now include part of 01/02, most of 02/03 and almost all of the rest. Coaches were included.

http://stats-for-the-nba.appspot.com/ranking
http://stats-for-the-nba.appspot.com/coaches

Even Michael Jordan gets a rating. Thibodeau is still #1 coach. Krystkowiak is dead last. Under him Milwaukee had a point differential of -7 with a lineup of Bogut, Redd, Villanueva, Desmond Mason and Mo Williams. Not the best lineup but -7? Phil Jackson got quite a boost through more years, I think. John Stockton looks incredibly impressive for a 39 year old.
If the Pistons replace Kuester with O'Brian this method would expect them to improve their point differential by +5.

LeBron and Garnett tower above everyone in the player ratings. Sam Cassell looks quite strong. According to this the future belongs to LaMarcus Aldridge, James Harden, DHoward, Chris Paul, LeBron, Amir Johnson to name a few.
Michael Olowokandi is dead last on offense with a -6.7. My guess is his teams were tanking. They had to be aware how bad he was and if they really wanted to win games they could have given him a lesser role. He was hovering around 19% usage though. Cracked 90 in ORtg once in his entire career.

Not accounting for age is still a problem. One I might tackle in the near future. If you're wondering why so many players have a negative rating: The ones with better ratings usually play more minutes.

The all NBA team for those years, according to this method, would be: LeBron Garnett Wade Ginobili Duncan With Thibodeau as coach
All defense: Garnett, Jason Collins, Artest, Doug Christie, Tony Allen coached by Scott Skiles
All offense: LeBron Kobe Wade Jamison DHoward coached by Tony DiLeo

J.E., again wonderful work. Would it be possible to add to the output how many possessions each player had for the ratings? That could be quite valuable in gauging the standard error for each player.

Also--I may attempt to rework my ASPM algorithm soon. Would it be possible to create a custom run for me, that includes regular season only, does not include coaches, and only includes seasons where you have nearly all of the data (otherwise the box scores for that season wouldn't match the data)?

DSMok1 wrote:Would it be possible to add to the output how many possessions each player had for the ratings?

Yeah I can do that

Would it be possible to create a custom run for me, that includes regular season only, does not include coaches, and only includes seasons where you have nearly all of the data (otherwise the box scores for that season wouldn't match the data)?

I can do that, too. It would probably be best to start in 03/04 then. Just be aware that there's not a single year with complete data. Not even '11.

You have to give me a couple of days for both of those things though.

On a different subject: Since I can crawl ESPN now, I can easily compute RAPM for the WNBA going back to 08. Does anyone know a website or forum where people are interested in that league?

Jerry, there's a SBNation log called Swish Appeal (http://www.swishappeal.com). I know one of the guys who runs it, because we both are bloggers for the Warriors. His name is Nate Parnham. I'm sure it would be a very valuable resource, as I am not aware of anyone else doing +/- for the WNBA.

How are you doing your crawling? What language? Just curious.

EvanZ wrote:Jerry, there's a SBNation log called Swish Appeal (http://www.swishappeal.com). I know one of the guys who runs it, because we both are bloggers for the Warriors. His name is Nate Parnham. I'm sure it would be a very valuable resource, as I am not aware of anyone else doing +/- for the WNBA.

Thank you.

How are you doing your crawling? What language? Just curious.

I use Python and urllib.
Each PBP is located at 'http://espn.go.com/nba/playbyplay?gameI ... '&period=0'.
Afterwards I just strip all the html with a bunch of "replace" commands and do some other things to make the PBP look more like bbv's

J.E. wrote:

DSMok1 wrote:Would it be possible to add to the output how many possessions each player had for the ratings?

Yeah I can do that

Would it be possible to create a custom run for me, that includes regular season only, does not include coaches, and only includes seasons where you have nearly all of the data (otherwise the box scores for that season wouldn't match the data)?

I can do that, too. It would probably be best to start in 03/04 then. Just be aware that there's not a single year with complete data. Not even '11.

You have to give me a couple of days for both of those things though.

Oh, there's no rush at all. It'll be a while before the next season starts! 04-11 gives me 8 years of data, and if you include the possessions played, it will be quite easy to weight the ASPM algorithm far better than my previous attempts. Thanks a lot, J.E.!

'03, '04 and '05 are now online. '06 and forward still needs to be updated with the new priors

http://stats-for-the-nba.appspot.com/ranking03
http://stats-for-the-nba.appspot.com/ranking04
http://stats-for-the-nba.appspot.com/ranking05

Garnett is always at the top. In '04 he's more than twice as good as the player ranked at #7.
Shaq and Duncan are usually at #2 and #3. Shaq more because of his offense, Duncan more because of his defense.
The next tier is usually Kidd, Rasheed and Dirk

J.E., I don't think it's exactly valid to regress to previous year if you don't adjust for aging. It will be strongly older-player-biased.

Thanks for the update!

Case in point: Andrew DeClercq in '05 played only 49 minutes and is in the top 15. That is exactly the issue RAPM was supposed to avoid!

Thinking about the Bayesian basis: I think the regression toward 0 needs to be of the same strength for each year (say, 2000 possessions), while the older year's data are weighted progressively less.

Right now, you weight the 0 on year 1 (correctly), then use that as prior for year 2 (but the year 1 data should be weighted less than the 0 weight). Do you see what I'm getting at? It would avoid the DeClercq problems with the current approach.

DSMok1 wrote:J.E., I don't think it's exactly valid to regress to previous year if you don't adjust for aging. It will be strongly older-player-biased.

I agree that there is a bias, but I don't think it is very strong.

Case in point: Andrew DeClercq in '05 played only 49 minutes and is in the top 15.

I don't care if some player ratings "don't look right". Can you prove that his rating is completely off? There is a good chance that his rating will be lower in a future, better version. But for now this version gives the best results

Thinking about the Bayesian basis: I think the regression toward 0 needs to be of the same strength for each year (say, 2000 possessions), while the older year's data are weighted progressively less.

The "2000 possessions" part confuses me. Multiyear analysis where I give less weight to older seasons is something I still haven't tried yet.
One year ratings with last years rating as prior outperforms zero as prior and outperforms multiyear analysis with equal weights to all seasons, so that's what I roll with until I find something better

I was just pointing out that you no longer are using a method that is unbiased or theoretically statistically appropriate. I know it's more accurate for most players (it's an informed prior!), but it is now biased towards older players, significantly. The more years you string together as priors, the worse it will get, if you don't reduce the weight of the older years while keeping the weight of the prior the same.