APBRmetrics

SUPER important thing to keep in mind while looking at the results that are questionable: the sample size. The authors mention that many of these players have under 20 games of SportVU tracking. Chris Paul had only 11 games! I think you guys are focusing too much on the results of the model and not the theory behind it. I don't think the authors meant to produce a definitive offensive value ranking. It's the first thing of this nature (and I think we can all agree that this sort of thing is the future of analytics) and there are growing pains.

It's the first step in the theoretical model we've all been dreaming of that gives a value to every movement on the floor.

Sethypooh, you make a great point, about the podium. Nobody in our group would ever say that Rubio is the worst player in the NBA or even that he is a bad player. In fact, we've had internal discussions about how heavily players' shooting abilities factor in to their overall EPVA scores - I think we weigh it way too much. Still, I saw an article in the Minneapolis paper with a headline that says, "Study says Rubio worst player in the NBA" - I hate that, but it's on me to figure out ways to not enable that kind of thing, and I'm not good enough "at media" yet. The David Lee thing from last year remains my biggest regret. You're exactly right about the conventional wisdom thing. ugh.

Jacob, thanks, the sample is both small and biased - this will be a big part of the presentation in a few weeks. But thanks for seeing the larger point about the first step in a long process. We have to start somewhere!

Everyone is being way too rough on EPV imo. It takes time to go from the original idea to the final version (of anything like this).

I do think that they pulled a McDowells (Coming to America reference) in regards to their naming of the metric

Last line of the Grantland article: "Sadly, as the best data sets become harder to acquire and the computational requirements become more intense, the days of bedroom analytics might be numbered."

Maybe, though that term appears to cast a fairly dismissive blanket over what has come before, short of super huge data analysis.
Efforts short of super huge data analysis still seem to offer unrealized value IMO but that is an outsider's guess.




It is and will be hard to know how much big data analysis is being utilized. It might be appropriate to measure how the 14 teams with access to sportsVU data previous to this year changed year to year on boxscore stats and how they changed relative to those that didn’t and how things changed from here and especially for the teams perceived to be the most involved with and best at big data analysis vs the rest.


With Prof. Goldsberry being an on-going Grantland contributor I am more hopeful that additional research will go public than I otherwise would be. Publishing future versions of EPV databases would be one way for ESPN to make a big contribution to the advanced analytics challenge.

sethypooh21 wrote:Thanks for responding Kirk. I think the biggest problem is only sort of in your control - given the podium you have, the "results" are going to be amplified and refracted across the basketball internet. Ricky Rubio worst offensive player in the league is now sort of a meme. I understand why you have to include examples in the article, but it still leaves a bad taste that these things quickly become conventional wisdom among a certain strata of observer.

This is what I feared because years after a Sloan paper, or even a small study on a blog, people will cite the results if the study is popular, even if the results are incomplete and out of date.

"the simplest method is to attribute them to the person in possession of the basketball at the time of those changes"

I have a question here. So what about cutting? Moving without the ball? Spacing? I understand the computational complexity, but this is a big part of basketball.`Will there be an attempt to incorporate this?

Kevin Love, for example, even using his games from this season, probably wouldn't do too well. He's great on offense because of how he moves without the ball and sets screens. He quickly pops out behind the line, where he's not significantly above average, so he wouldn't get a lot of credit here because he'd be compared to the "average" player.

Right now, from the EPV model I've seen, it appears it would do well at rating ball-handlers but not as an "all-in-one" metric. But it seems like something that'd be great at rating defenders because it's all about causing actions that increase/decrease expected points. (Schemes/mismatches might be tricky though.)

Okay, so, I don't expect you to take any advice from me, of course (my opinion of myself isn't that high), but I think it would be better to tackle things one at a time. Like starting with a metric only looking at passing/ball-handling and rating them based on how they find high percentage opportunities for their teammates and putting their teams in a position with a higher expected outcome than at the beginning of the possession. Or look at screen/picks, because I feel that's one of the most overlooked aspects in the NBA, even with all the data we have now. Just a thought.

I do agree with some of the fears in this topic about this study being released into the public when it's not complete, causing a lot of people who don't understand how this is a "beta" version with a small set of games to parrot the results to claim certain players are under/overrated.

And we didn't mean to be so harsh on the work here. This just quickly became a small gathering of people to talk about a few negative things with the paper.... It is a really exciting area and I was fascinated while reading it, but when looking through this topic I wanted to share my criticism.

The EPV framework is fantastic! Props to Dan, Alex, Luke and Kirk.

That doesn't mean that you assign the ball handler all of the credit for every action, when all of the actions, as well as the values of those actions, depend on the other players on the floor in the first place. The player values therefore aren't valid and aren't nearly as important as the EPV framework itself. It'll take a while to evaluate one player's total (on-ball and off-ball) value within the context of the other players on the court, but this is a good start.

Where the EPV framework will really help right now is for play calling and in-game decision making, much like Rucker and Boyarsky's Raptors SportVU algorithms. It's tremendously useful to know that certain spacings on the court (for offense and defense) are much more valuable than others depending on personnel.

And this is still the problem with bottom-up models that assign values to discrete actions: it's unclear how to distribute credit between the Tony Parkers and the Kawhi Leonards. Until there's a way around this, top-down APM type things will be more accurate.

Jacob Frankel wrote:And this is still the problem with bottom-up models that assign values to discrete actions: it's unclear how to distribute credit between the Tony Parkers and the Kawhi Leonards. Until there's a way around this, top-down APM type things will be more accurate.

Or, perhaps, you use APM results to figure out how to assign those values.

DSMok1 wrote:Or, perhaps, you use APM results to figure out how to assign those values.

That's an excellent idea. Perhaps even more logically, assign it a delta win-percent (or delta-win-probability weighted Points) since we're already in that sort of frame of reference. This would prevent scrubs from getting tons of credit when they aren't affecting the win.




Edit: Now I'm daydreaming about the extreme importance of cross-validation in a clutch/garbage time-adjustment....

Thanks, everybody - there is some great stuff in here. Off-ball stuff is key and dividing credit is also really important. I also love the idea of weighting by win probability. Hope to see you at the Fours.

What if instead of calculating EPV moving forward thru a possession, you did it backwards and maybe in just three parts. The shooter / scorer gets credit / blame for use of the possession except for the portion where EPV at time of pass reception is above league average. That portion is divided between the last passer and everybody with the last passer getting say 25-50% of the difference and everybody getting a share of the remainder in proportion to their (ORPM- off.reb portion of ORPM) - (OBPM-off. reb. portion of OBPM) and the cumulative sum of this value for the 5 players on the court, to capture their average rate of indirect contributions to offense, so as to avoid having to subjectively value all the pieces of preliminary movement, spacing including picks and passing on every unique play.