APBRmetrics

[Mountain]

I increased the labeling; it was insufficent on those tables previously, sorry. I provided links and brief descriptions on the better links above including the study that provides those tables and initially was going to just leave it at thumbnail description and let any interested explore / react but decided to pull those tables out to perhaps pique interest in reading and talking about the BYU study of position/ boxscore contributions to team success.

Basically the posterior means are intended to measure importance (or predictive power of) various skills at each position, as represented by boxscore stats, in future team success using Bayesian techniques.

But for a full description, read the article. I don't any detailed notes on their methodology because evaluating it is more in the province of others with better stat theory backgrounds. I got some basic messages from their tables that seem straight forward after reading the summary pages but rather than going into what I got out of it at length immediately, I decided to wait and see if it interested anybody else first. I was hoping someone with a stronger stat background would assess the quality of the work before I also trumpeted the conclusions too much and I thought they could probably also do a better job of summarizing what was done and what exactly it meant than I could. Posterior means included but I hope I described it acceptably.

I probably presented too many links earlier so I adjusted that a bit too. I'd once again highlight the Casell shooting study and note that the Lucy Liu abstract looking at "social relations", player contributions to team success sounds like it might be worthy of attention, especially by teams and their professionals.

Here is the Liu study link

http://www.stat.sfu.ca/people/alumni/Theses/Liu-2007.pdf
It uses the posterior means term to describe its team impact results as well as it is also using Bayesian techniques (and mentions Markov chains along the way and compares result sto raw team +/-). "The author spent close to 240 hours transcribing 12 games of videotape into the required data format." Impressive effort though detail about the nature of social relations or uncounted player contributions isn't what I'd hoped for in terms of defining these contributions or giving player scores. Is there more detail beyond the paper?

Data taken from 04 and 05 NBA finals. Check it out. Ben Wallace the leader in 04 Finals but Rasheed Wallace and Ginobili led 05 results. I'd be interested in know what the strong statistical analysts think of it. Sent e-mail to the thesis advisor inviting them to discuss the work here if they wish.

[KShirley]

Hi, I just signed up for the message board - so far it's been very exciting to read. I think Mike summarized my Markov model from NESSIS quite well in his earlier post, so for now I just thought I'd add the link to the poster and make a few general comments:

Poster Link: http://www-stat.wharton.upenn.edu/~kshirley/kshirley_nessis07.pdf

A few comments:
1. I've fit two models so far, with different state spaces and different sources of data.

The first one was an 18-state Markov chain whose transition probabilities were estimated from season-long team-level stats like 2pt FG%, offensive rebound %, turnover rate, etc. This model made certain assumptions like no 4-point plays, and no offensive rebounds on missed free throws. It made other, more important and harder to believe assumptions about the frequency of deflections and the increased FG% that results from steals and offensive boards. Still, this model fit pretty well, and the winning percentages of the 30 teams in the NBA were well-estimated via simulations from this model. This is the model that I called a "pilot study" on the poster - i don't think it's really worth studying further - it just convinced me that a Markov model might be a good model for basketball.

The second model is a 30-state MC whose transition probabilities I estimated by recording actual transition data by watching 4.5 games on tape - too small a sample to really make any conclusions, but enough, i hope, to give a glimpse of the usefulness of the model. This larger model requires no assumptions, as it contains enough states to account for very rare events, like the unusual things that can happen on missed free throws, and the sometimes irregular possession patterns that result from flagrant and technical fouls. I think this model is interesting and could shed light on the value of offensive and defensive rebounds, as well as the value of defensive deflections. But to estimate these things with any precision, much more data is needed. I'm thinking about watching a bunch more games, hiring some people to do that, or trying to find a play-by-play source of data that includes all of the things my model requires (deflections, for example, don't show up in many places).

I'd love to hear any more comments or suggestions from members of the forum!

best,
kenny

[Mountain]

I wonder if Lucy Liu's dataset mentioned above would allow you to work your model with at least a little more data. Her study talks about Markov models but I am not equipped to immediately see how similar it was to your work. Tim Swartz her advisor did contact me back and said he would have done some things differently, thinks there is more to be gotten out of such an effort. Maybe you and he might have a good conversation? Perhaps Liu's videotape work, if it is satisfactory to your needs, on 12 games might help? From your poster I see now that ideally you want a lot more data. Maybe it could be a version 2 for further model refinement rather than the ultimate version. Maybe with help of more grad students (or others) you can add to it?

Bio and contact info here
http://csc.sfu.ca/bio.php?id=455

[HeatherA]

Kenny,

Welcome! I believe that we spoke at the Fours about video coding issues? Here is a page with the contact information for Synergy:

http://www.synergysportstech.com/website/extranet/Login.aspx

They have coded games to an amazing degree of specificity. They charge teams big bucks for access, but you might be able to work out something with them to get access for research purposes.


On another note, what software did you use to do your poster for the conference? I'm preparing my first conference poster in a long time and would love to move beyond the scissors and construction paper stage.

[Mountain]

It is possible that protrade.com might have some useful data. it was in 05-06 I believe that they computed expected value of play by play actions toward winning on some basis, though they no longer show such data publically having gone to a fantasy scoring basis. It was for boxscore actions but perhaps there was unmined detail. Not sure if the data is still around or if they'd share it but it might be another lead that could be checked if other sources don't fulfill the need.

Or perhaps Roland Beech at 82games (using his play by play dataset) or even an NBA team (using their data off video) might be interested in some collaborative arrangement. If they collect all the elements you want or would be willing to modify to do so this season.

[KShirley]

Thanks for the link to Tim Swarz's site - maybe I'll get in touch. For now it looks like Liu's work uses Markov Chain Monte Carlo (MCMC) methods for fitting her model, but the model itself is not a Markov Chain model - it's a generalized linear model which is pretty similar to +/- ratings, where the raw data consists of intervals of time during which the same 10 players shared the court.

Heather, thanks for the synergy link - i'll check it out. My poster was done on microsoft power point; i just manually adjusted the slide size to be 36'' by 56'' (somewhere in one of the menus this can be done easily), and then inserted text boxes and figures and moved them around the page to try to make it look official

[Mountain]

Alright. I thought initially that Liu's study might show greater detail on non-boxscore actions and their value. But as you say the results don't show that and rating players is different than your goal. But worth a further mention I thought.

The BYU study, also referenced above, was just boxscore actions (and position) but perhaps the output, value of those play actions, might be of some interest.

Thanks for sharing the poster and good luck with the next steps.

[Eli W]

For those who weren't able to attend the conference, there are now videos available of the presentations:

http://www.amstat.org/chapters/boston/nessis07/videos.html
_________________
Eli W. (formerly John Quincy)
CountTheBasket.com

[Mountain]

Thanks for the heads up. This is a generous public outreach step.

[tpryan]

Yes, thanks for posting that link. I've known Carl Morris and Cyrus Mehta for many years. Enjoyed hearing and seeing them.

[mtamada]