The way basketball defense statistics can get better

[sicknastydunker]

Baseball defense statistics have gotten so much better because someone(MGL) bought hit location data from BIS and made a great defensive stat out of it. The way to do this in basketball is either synergy to release data to the public(not going to happen) or more people to collect PDSS data. PDSS is very hard to collect. DRAPM is a good placeholder, but not the actual data. The way I am thinking is that Sportsvision uses their cameras for basketball defense. Sportsvision made an incredible Pitch F/x system for baseball that tracks every pitch's location, movement and speed. They have also made a hit f/x which tells you exactly how hard the ball was hit and the location. They also made a field f/x which records everything from hang time of batted balls to hit location which is better than BIS data to baserunning. Unfortunately hit and field f/x haven't been released to the public. But if sportsvision created one of these systems for basketball, defense would be so easy. The way to do it would be to track every player's movement. We could then even measure help defense. I hope Sportsvision does create this, and hopefully the NBA teams release the data to the public.

[Doctor MJ]

Well I actually think it's less difficult in some ways but also less straight forward than that. Baseball is so clear how to track because it's essentially a hybrid between a sport and a board game with every piece in a given position. To really get it all down then, super-precision in the automated tracking is really a reasonable expectation.

In basketball, when judging defense, there's less obvious benefit for judging shot location, and far more benefit in classifying situations, actions, and consequences, which means that super-sophisticated technology isn't as important as intelligent manpower. Key word is "intelligent". There's simply no easy way to program a computer to properly place blame in many situations (it being unacceptable to program a computer to associate "pick & roll" + "Amare defending" with "Amare costs his team 2 points" in all situations).

[Chicago76]

Agree with Doc completely. Baseball is a series of fairly independent discrete events. It's easy to measure pitch location and speed because you know where the pitch is going to be +/- a couple feet. It's also easy to measure trajectory of a batted ball, because it is easy to measure: 1) speed of ball coming off the bat 2) location of the hit with less precision required +/- a foot or so in baseball is generally sufficient, compared to basketball and 3) the time it took from contact with bat to the time it hit the ground or a fielder's glove. That's basically all you need to know to compute trajectory, spin, etc.

Let's take a couple relatively complicated baseball plays. Man on first with one out. Line drive hit down the right field line. How good is the right fielder at either throwing the man on first out at third or holding him to second? You have the trajectory data and you know how long it took the RF to retrieve the ball. You know how far he is throwing from and how fast the runner is (based upon 162 games of data). You can use this to compare the RFers success rate compared to other times the same runner has been in this position (+/- a few feet on the throw, which isn't too significant when you're talking abt 300 feet or so). It's the same thing with measuring a SS-2B double play combo given certain runners.

Basically discrete plays are easier to data mine, decisions on what to do in those plays are less subjective (based upon base runners, score, inning, and fielder), and you have 162 games to mine the data.

In basketball, you've got 10 guys on the court moving somewhat independently. You've got a smaller court, so inches matter more. Technically, it is much more difficult to measure the flight of the ball and player movement because the play can happen virtually anywhere on the court. How you defend will be hugely dependent upon the shot clock. It will be even more dependent upon the personnel you have as teammates on the court and the strengths and weaknesses of individual opponents. Example: does a wing defender help on dribble penetration? It depends upon how good the defender of the penetrator is. It depends upon how good the opponent is at penetrating and passing. It depends upon how good the post defender is. It depends upon how good of a shooter the wing defender will need to cheat away from. And so on. Each player is making dozens of mental calculations each second in basketball. In each baseball play, a fielder is making 2 or 3 calculations before a play happens. By the time a ball is actually hit, the decision has already been made.

Measuring defense in baseball is a lot like tracking moves in a board game like Sorry. Measuring defense in basketball would be a lot like tracking moves in backgammon...if backgammon was played 3 dimensionally...and 6 players were rolling the dice and moving around the board simultaneously instead of waiting their turn.

[wiLQ]

But if sportsvision created one of these systems for basketball, defense would be so easy. The way to do it would be to track every player's movement. We could then even measure help defense. I hope Sportsvision does create this, and hopefully the NBA teams release the data to the public.

IMHO the answer to your question is simply "we need more data".
What you've described sounds to me like a best-case scenario in a perfect world but I doubt someone would put so much money and technology without selling it to teams so we should probably hope for more basic columns like "deflections", "switches", "blowbys" etc.

[sicknastydunker]

But if sportsvision created one of these systems for basketball, defense would be so easy. The way to do it would be to track every player's movement. We could then even measure help defense. I hope Sportsvision does create this, and hopefully the NBA teams release the data to the public.

IMHO the answer to your question is simply "we need more data".
What you've described sounds to me like a best-case scenario in a perfect world but I doubt someone would put so much money and technology without selling it to teams so we should probably hope for more basic columns like "deflections", "switches", "blowbys" etc.

What I was thinking was that they would sell it to the NBA teams and some company like MLBAM would want to put it on a gameday like app and then publish all the data from the gameday like app on the site, that's what happened with pitch f/x. You can get the raw data here under innings.http://gd2.mlb.com/components/game/mlb/year_2011/

Well I actually think it's less difficult in some ways but also less straight forward than that. Baseball is so clear how to track because it's essentially a hybrid between a sport and a board game with every piece in a given position. To really get it all down then, super-precision in the automated tracking is really a reasonable expectation.

In basketball, when judging defense, there's less obvious benefit for judging shot location, and far more benefit in classifying situations, actions, and consequences, which means that super-sophisticated technology isn't as important as intelligent manpower. Key word is "intelligent". There's simply no easy way to program a computer to properly place blame in many situations (it being unacceptable to program a computer to associate "pick & roll" + "Amare defending" with "Amare costs his team 2 points" in all situations).

But my idea was that sportsvision could track the movement and location of all the players on the floor. If we knew where each player was when the ball was shot, we could use a LOESS regression curve to find out what percentage of players in that location could have helped. That's just one thing. If we had this data, the defensive statistics would be so good, but of course we don't, and even if sportsvision makes this, it probably won't get released to the public, but it's just wishful thinking.

[Chicago76]

Even knowing floor location won't tell you much from raw data only. For example, if a defending SF is 6 feet away from D Wade offering a bit of help, does he get blamed for a Wade basket? He might have cut off a penetration angle while not entirely losing his primary responsibility: James.

You'd need to know distance between all players, which way everyone is facing, which way everyone is moving, how fast they are all moving, the primary defensive responsibility of each player, and you'd need to weigh the ability to help vs. the need (for every step you move away from your primary assignment to help, how much greater is the threat of your primary assignment vs. how much do you diminish the ball handler threat?)

Defensively, everything is a matter of inches with a ridiculous number of variables, so what you're able to measure would be unique to the nth degree, making it difficult to get the number of observations to draw major conclusions.

Teams are using this type of stuff now, but it's still pretty general: how players defend a pick and roll, how they can force a player into a lane where help is available, etc. It's more film study and being able to call up similar plays on video to detect trends than it is an uber complex set of measurements.

[Doctor MJ]

I agree with Chicago76, and I guess my gist is this: While I'm certainly not against this type of data, it's not the #1 thing on my wish list. What I really want is more advanced tracking by human observers. Now, you might say "automated is cheaper than humans", but in this case, maybe not by as much as you'd think. Basketball is the rare sport where you can typically see or infer all the action that takes place simply from video, and there is brainpower to spare on the internet.

[Nivek]

Advanced tracking by human observers is the ideal way to do this. I've done this myself for a single team and it's completely feasible and yields robust information. A couple years ago (when I was working on a tracking proposal for a team), I did a back-of-the-spreadsheet calculation that every game in the NBA season could be tracked for something around $1 million per season -- not including start-up costs, which wouldn't be much. That million included salaries for 15-16 full-time trackers, a programmer and a supervisor/quality-control person.

[SinJackal]

Not to throw a wrench in here, but you also have to measure a player's drop in FGAs into the mix somehow too. If you deny a great offensive player the ball with excellent off the ball defense and positioning, no defensive stat currently reflects that. You can say well look, he shot 60% today. . .but if he only scored 12-16 because he rarely got the ball (while typically averaging say 25 a game), that's a great defensive night. The defender effectively cut the star's impact on the game in half, but the stats don't reflect it sadly.

Another reason I mention that is because let's take a player such as Kobe Bryant for example. If the Lakers play. . .hmmm, the Heat we'll say. Kobe doesn't guard Wade, he'll guard Chalmers. . .a guy who barely takes any shots and doesn't handle the ball much. If he holds his guy to 35% and 2 assists. . .it doesn't mean he was defending great. The player he was guarding just isn't very good. Or if they play the Thunder, he won't help Fisher out and guard Westbrook, he'll guard Sefolosha, who barely takes a shot. Ot if they play the Mavs, he doesn't guard Marion or Terry, they switch Fisher to those guys and Kobe gets to take a rest by guarding Jason Kidd who barely even shoots the ball anymore.

So his defensive stats would look great on paper. . .since they don't mention the players they're guarding. But a guy who does that on D Wade, LeBron, Durant, etc. . .his stats wouldn't look as impressive on paper, but in reality he had a much larger impact on the game by far if he held those guys down, as opposed to just locking up borderline scrubs like Sefolosha and Keith Bogans.

Still, I do hope defensive stats at some point start showing how much impact these players can really have. They don't get much love unless they get a lot of steals or blocks.

[EvanZ]

I've been running a series of adjusted shot analyses as an orthogonal metric for offensive and defensive evaluation. I think it's relevant to this discussion and specifically, addresses the SinJackal's problems with most defensive metrics. Brief explanation:

In the previous post, I did a WOWY (with-or-without-you) analysis of high-volume mid-range shooters and their affect on the team-level shots. I pointed out at the end of that article that it would be better to turn this into an adjusted metric using linear regression similar to the way APM or RAPM is calculated. Here, I’m going to take a crack at it using ordinary least squares regression on data from 2010-11.

To do this, I used the same basic methodology as APM. I counted mid-range shot attempts by the home and away team for each “stint” during a game. A stint is a series of possessions that occur with no personnel changes. However, instead of having one factor for each player, there are two factors, one each for offense and defense. There is also a factor for homecourt advantage (HCA). I used 1000 possessions as my minimum threshold. And in the following spreadsheets, I only give coefficients for players that were found statistically significant (p<0.05). It turns out that the standard errors (SE) are actually quite small compared to 1-yr APM, for example.

1-yr Adjusted mid-range shot attempts

3-yr adjusted mid-range shooting

3-yr adjusted inside+FTA rates

3-yr adjusted mid-range efficiency

3-yr adjusted inside shooting efficiency

Coming next is adjusted 3-pt shooting.