Trying to Understand Every Move They Make: EPVA Stat

[slaw]

Fascinating stuff from Kirk Goldsberry at Grantland.

Every “state” of a basketball possession has a value. This value is based on the probability of a made basket occurring, and is equal to the total number of expected points that will result from that possession. While the average NBA possession is worth close to one point, that exact value of expected points fluctuates moment to moment, and these fluctuations depend on what’s happening on the floor.

In other words, imagine if you paused any NBA game at any random moment. Cervone and D’Amour’s central thesis is that no matter where you pause the game, that you could scientifically estimate the “expected possession value,” or EPV, of that possession at that time.

[I]If we can estimate the EPV of any moment of any given game, we can start to quantify performance in a more sophisticated way. We can derive the “value” of things like entry passes, dribble drives, and double-teams. We can more accurately quantify which pick-and-roll defenses work best against certain teams and players. By extracting and analyzing the game’s elementary acts, we can isolate which little pieces of basketball strategy are more or less effective, and which players are best at executing them

http://grantland.com/features/expect...nba-analytics/

[OldSkoolCool]

The Parker/Leonard play they broke down was pretty sweet.

They have a long ways to go yet but the EPV is a massive step over using box scores

[Axel]

Don't know how I missed this post. Fantastic read. Love the Goldsberry work this season on Grantland.

[stooley]

I read this article too. Really interesting. I was thinking about win shares the other day and how even basic synergy stats like rim protection, and shot charts could really improve it.

I found this post from real gm that summed up the stat really nicely:

From chicago76, http://forums.realgm.com/boards/view...44&p=29891896:

The resources on the net do a pretty inadequate job of explaining Ortg, Drtg, and WS and Dean Oliver's chapters on it in his book are pretty complex, so it looks like a very simplified nuts and bolts explanation is in order. There are several ways to calculate WS, but they're all variation on what is essentially the same theme. This will only be 95% correct, but it's the 95% you need to understand indiv. Ortg, Drtg, and WS.

The first thing to understand is that WS does not attempt to tell you how many wins a player is responsible for. What it actually does is attempt to allocate credit for wins a team should have achieved, given its offensive and defensive efficiency (or pt differential, since possessions for both teams are roughly equal over a season anyway). This is a pretty small distinction, but an important one, because a team's Ortg and Drtg will set a benchmark for each player.

We get expected team wins from the difference between off and def efficiency in the following formula: Ortg^14/(Ortg^14+Drtg^14) x total games. Assuming team Ortg = 108, Drtg=102, a team would be expected to win 57 games. If league avg. efficiency = 104, then 31 wins come in the form of OWS and 26 come in the form of DWS because the offense outperforms the league avg more than the defense does. Note: you might think the ratio should be 2:1 Off wins to Def wins, but it isn't for a reason we don't need to get into.

These two pools of OWS and DWS (and the corresponding Ortg and Drtg) from the example above are the parameters used to determine individual WS for a player. Individually:

Ortg: is a linear weights measure which tabulates the positive things a player does (made FG, FT, AST, and ORB) against the negative (TOV, missed FG, missed FT) that can be readily obtained in a box score. The objective is to quantify points generated and possessions used for each player to arrive at Ortg for each player. If a player nets to exactly what the team does per poss, then their rating should also be 108. Do less, and you're less than 108. Do more, and you're over 108. Each player is anchored somewhat to the team in such a way that if you possession-weight everyone on a team, this should reconcile to the team Ortg of 108 in our example.

Drtg: a linear weights measure which tabulates all of the positive defensive things a player can do that is captured in the box score: STL, BLK, and DRB. Missed opp FGs and turnovers not from steals are evenly distributed based upon minutes, which is part of the credit problem we'll get into later. Everyone gets a Drtg based upon these "stops" per poss. Just like the offense, if you do more, your rating will be better and so on.

You can then use the formula for team expected wins above on an individual level. You simply need to apply indiv. Ortg and Drtg into the formula just as you would for the team. OWS and DWS are possession-weighted and minutes weighted, respectively vs. team totals to get to this score.

Before we get into the credit issue, it's also important to note that even if a player performs below team avg, some credit will be given. As an example, let's say a player performs 10 pts worse than league average offensively and plays the entire season. On a team level, if a team is 10 pts worse than league average on both ends of the court, they're all-time awful. The would be expected to win 5 games (2.5 on each end of the court) out of 82. So a player responsible for 1/5 of the team's possessions who is 10 pts worse than league avg offensively is still worth 0.5 wins on that end of the court.

This nice thing about WS is that it reconciles well with the team. Now for the bad things, using the player examples earlier in the thread:

-Olajuwon. The value of elite interior defenders is undervalued due to the defensive credit problems. Olajuwon produces value with BLKs, STLs, and DRBs, but a big part of his value is in missed FGs, which are evenly credited among his teammates based upon minutes played. Opposing offenses don't even enter the lane to even attempt shots vs. Olajuwon. Instead, they rely upon lower percentage mid-range and perimeter shots, and Olajuwon doesn't get credit for this. The Rockets went from being the second best D in the league in 93-94 to an avg team (12th) in 94-95. Why? The 3 pt line moved in and their perimeter D wasn't particularly good. Those 2s teams were taking before to avoid Olajuwon were now worth 3 pts. Teams took more of them, and there were longer rebounds on those misses which more frequently became ORBs. On the offensive end, 3s became a more efficient weapon too, making Olajuwon's offensive production relatively less efficient. Olajuwon wasn't worse. The rules changes just made him relatively less valuable. The Rockets adapated their D in subsequent seasons and improved here (along w/ Olajuwon's DWS). Also: in 94-95, Hakeem only played 72 games vs. full seasons for Schrempf, Miller, Barros, which certainly hurt his gross WS number.

-Kidd. True creators are undervalued in WS. They provide easy shots for teammates, and the incremental production is credited in the form of efficiency to the scorer. Sure, there is an assist component, but all assists aren't equal. A guy who can demand double teams on penetration before dishing has value that a more run of the mill PG doesn't have. The same goes for scorers. Speaking of which, Kidd was horribly inefficient as a scorer. This killed his Ortg and WS. Also: great defenders who don't necessarily generate a ton of steals or blocks will be undervalued. Kidd did get a lot of steals but hardly any blocked shots.

-Very efficient third and below option scorers are almost always overrated. These players are efficient because they're good at knocking down shots the defense concedes to take away the primary scoring threat. In a perfect world, some of this credit would be given to the lead offensive players, but how much? There is no clean way to allocate. This is why the Jon Barry's of the world (and Detlef Schrempfs on more balanced teams) can generate huge WS. The distinction here isn't so much efficiency as where in the offensive pecking order a player is. Steve Kerr got wide open looks and was efficient in hitting them thanks to Jordan, Pippen, Kukoc, and the need to defend easy buckets inside. He was the net beneficiary of great scorers/creators. Reggie Miller on the other hand was deadly efficient as a primary scorer. He played off Smits probably about as much as Smits played off him, but defenders weren't exactly preoccupied with the Davises, Derrick McKey, the shooting touch of Mark Jackson, etc. Miller's deep threat at volume benefitted team spacing and offense rather than the offense benefitting Miller.

Hopefully this helps someone conceptually. The stat isn't flawed. It's a gross stat (as opposed to PER, which is a rate stat). As far as gross stats go, it is a very useful one. Just compare WS to another gross stat (like total points scored in a season) and tell me which one does a better job of ranking players. Like any statistic, you just need to understand what it can and can not measure rather than discarding it indiscriminately.

[Axel]

This is from the research paper referenced in the original article. Some interesting names for Raps fans. This is based on last season's data.


Top 10 and bottom 10 players by EPV-added (EPVA) in 2012-13 (per game, minimum 100 touches during season). Because of limited SportVU coverage in the 2012- 13 season, some players in these tables were only observed for a small number of games despite having enough touches to be included. For example, Chris Paul (ranked 1st) and LeBron James (ranked 23rd, not shown) were only observed in away games, for 11 and 17 games, respectively. In the se games Paul shot 54 % compared to his season average 48%, while James shot for 50% in our data compares to his season average 57%. This sampling bias accounts for some anomalies in this ranking, which could be eliminated by considering the 2013-14 season's complete data.

Top 10
Paul 3.48
Dirk 2.6
Deron 2.52
Curry 2.5
Crawford 2.5
Vasquez 2.46
Aldridge 2.4
Nash 2.09
Matthews 2.06
Lillard 1.95

Bottom 10
Rubio -3.33
Love -2.38
Westbrook -2.07
Turner -1.9
Rivers -1.84
Gay -1.75
Holiday -1.51
George -1.49
Singleton -1.48
Hibbert -1.44

[stooley]

Top 10
Paul 3.48
Dirk 2.6
Deron 2.52
Curry 2.5
Crawford 2.5
Vasquez 2.46
Aldridge 2.4
Nash 2.09
Matthews 2.06
Lillard 1.95

Bottom 10
Rubio -3.33
Love -2.38
Westbrook -2.07
Turner -1.9
Rivers -1.84
Gay -1.75
Holiday -1.51
George -1.49
Singleton -1.48
Hibbert -1.44

Interesting that there are so many point guards in the top 10. Makes sense.
The presence of Aldridge and Dirk are interesting too. Are they mid-range point forwards?

Love, Westbrook and George are all surprising to see in the bottom 10. Any ideas as to why those guys show up?

[Axel]

Interesting that there are so many point guards in the top 10. Makes sense.
The presence of Aldridge and Dirk are interesting too. Are they mid-range point forwards?

Love, Westbrook and George are all surprising to see in the bottom 10. Any ideas as to why those guys show up?

PG definitely makes sense since I'm pretty sure the data on calculates for a player while they are in possession of the ball. So Duncan setting the screen doesn't affect Duncan's stats, it does impact the expected value for the players around him (the drive, the defence rotating off a guy, etc) that Parker then has to make the decision from. Dirk and LA both hold the ball in a high post a lot for their offences, which gives them plenty of options to score, pass, drive, etc.

I'd wager Love scores so badly partially due to playing through his hand/wrist injury last year. No idea about the others. I was more surprised to see both George and Hibbert there since they were both starters on a very good Pacers team that doesn't have a true PG.

[stooley]

I'd wager Love scores so badly partially due to playing through his hand/wrist injury last year. No idea about the others. I was more surprised to see both George and Hibbert there since they were both starters on a very good Pacers team that doesn't have a true PG.

Good point on Love, didn't think of that. I'm not surprised that Hibbert scores badly though. He's not known for his offensive game, and neither is the rest of his team.

[Axel]

Good point on Love, didn't think of that. I'm not surprised that Hibbert scores badly though. He's not known for his offensive game, and neither is the rest of his team.

I took the stat to be more about making the right decision, but as Hibbert doesn't likely make too many passes, I guess it's not that surprisingly. Also, as the preamble states, the stats are very limited in sample size since only a few teams actually had the Tracker system in place last year. The stats from this season will be much more indicative.

[Uncle_Si]

I've got to give credit where credit is due. There are some pretty sophisticated basketball minds on these forums. From more of just 'fans' perspective it's fascinating to read some of these discussions. Cudos to you all. Keep it up.

Also if you're willing to answer I'd be interested to know how many of you have a statistics background of some sort and how many are just nba state nerds

[slaw]

I took the stat to be more about making the right decision, but as Hibbert doesn't likely make too many passes, I guess it's not that surprisingly. Also, as the preamble states, the stats are very limited in sample size since only a few teams actually had the Tracker system in place last year. The stats from this season will be much more indicative.

Yes.

It should be noted that this only addresses actions with the ball. The next steps, and the ultimate idea here, is to try and determine what influences a player has when he doesn't have the ball in his hands. In other words, trying to ascribe an actual value to a screen or a hedge.

Way beyond me in terms of the math involved (and beyond my limited time to try and figure out) but fascinating. Interesting to see if it goes anywhere useful....

[Superjudge]

Perhaps they should just assess everyones game through a myriad of intense and rigorous set of tests. They could take the results of these teats, and using special algorithms, they could put all the players on a team together and calculate the maximum potential. Then they could compare how they match up in each of the potential 82 games they would play, and determine their win loss...


then

You wouldn't have human beings to blame for the game swe watch, and, we wouldn't even need to play the games.

[stooley]

You wouldn't have human beings to blame for the game swe watch, and, we wouldn't even need to play the games.

Stats don't kill the game baseball still has hero moments and clutch shooters. It just adds an element beyond the media's narrative.

[thead]

Is there an example of a heavily analytics influenced draft pick? I just feel like well number crunching can be fun and I appreciate a lot of the stats that value good shots over bad shots...a lot of it is common sense.

Did we need advanced stats to tell us that 19ft jump shot with a hand in your face is statistically worse then a wide open corner 3, or did we all kinda know that already.

Do we need stats to tell us guys like Steve Nash, and Ricky Rubio make the game easier for their team mates or can we just clearly see that watching the game.

I can see that Hibbert is an excellent rim defender by watching the game.

Has anyone changed their opinion of a player based on an advanced stat?

[Axel]

Is there an example of a heavily analytics influenced draft pick? I just feel like well number crunching can be fun and I appreciate a lot of the stats that value good shots over bad shots...a lot of it is common sense.

Did we need advanced stats to tell us that 19ft jump shot with a hand in your face is statistically worse then a wide open corner 3, or did we all kinda know that already.

Do we need stats to tell us guys like Steve Nash, and Ricky Rubio make the game easier for their team mates or can we just clearly see that watching the game.

I can see that Hibbert is an excellent rim defender by watching the game.

Has anyone changed their opinion of a player based on an advanced stat?

Funny that you mention Rubio when he actually scored terribly. After the reading the article, I started to wonder if Rubio has what it takes to be a legit starting PG. The biggest statistically downfall for him in this case is he awful shooting; it's bad enough to negate his passing and vision benefits compared to the average player (I'd be very interested where exactly the average lies and which player is closest).

I think the value here is best demonstrated by the Parker-Leonard play they use in the article. Parker gets credited for an assist but doesn't currently get awarded for making a best decision. If that exact play happened and the names were Steve Blake - Darius Morris and it was done midway through the 2nd quarter; would it get the same level of attention? Not likely; so the stat can show "hidden " value. Overall though Thead is right in the sense that it backs up the eye test but the true value is that is covers all players all the time, which unless your job is covering the NBA and you have lots of time, there is no way to get the same coverage with the eye test.