NBA-analytics

Short, offhand analyses of the NBA

Note: All stats are from stats.nba.com, unless noted.

Topics Covered

Underrated Assisters
Consistent Players
Why is Miami the worst team at drives?
Which teams run the wrong plays?
Shot Clock Analysis
Westbrook Rebounding
Defender distance
Aging player's efficiency
Steal efficiency
Timeout Conversion
Rim Protection
Tanking
Empirical Bayes estimation of NBA statistics
James-Stein estimation of NBA statistics
Player Efficiency Rating (PER)

Underrated Assisters (December 2016)

I wanted to look into underrated assisters in the NBA. Players who are really good at creating plays by passing, but don't necessarily rake up traditional "Assists" in their box score. I thought I would find players who are thought to be bad/average, but are actually elite. What I ended up finding is that a few elite players are even more elite than we think.

First, are there players who don't get a ton of traditional assists, but get a lots of "Secondary Assists"? The guy who passes it to the assister. The hockey assist. The answer is Steph Curry. Below, I'm plotting the total primary and secondary assists of every player in the NBA this season.

In fact, the Warriors are a standout team when it comes to total hockey assists. It's not even close.

Another thing I wanted to look at was if there are players who's assists are more valuable than other players. If a player assists a lot of 3PT shots, that is more valuable than a lot of 2PT assists. Again, it turns out that two all-stars, LeBron James and James Harden are at the top of this list. My favorite anomaly in this category is Harrison Barnes who has 34 3PT Assists and 3 2PT Assists. He exclusively assists to 3PT shooters.

Finally, I thought I could find underrated assisters by looking at "Potential Assists". A Potential Assist is a pass that leads to a shot, even if the shot does not go in. I thougt this could be used to find players who are good as passing, but they are on bad teams and the shooter never makes the shot. I turned out to be wrong: All players in the NBA essentially have a 2:1 ratio of potential assists to assists.

Consistent Players (April 2017)

Analyst typically say things like "Player X brings it every night" or "Player X will give you 15, 5, and 5 every night". Here we try to quantify how consistent players are. The conclusion is that LeBron James is a very consistent player.

Here, we use game logs from the 2016-17 season. To get a sense of how consistent a player is, we take the standard deviation of a statistic over all games in the season. A lower standard deviation is indicative of a more consistent player.

Who are the most consistent scorers? Here we see that LeBron James is one of the most consistent scorers in the league. He scores over 25 PTS/game, and does so very consistently. Other players are more consistent, but score much less points than him.

What about getting to the free throw line? Again, LeBron is an outlier. He is very consistent at getting to the free throw line.

Same with rebounds and steals. LeBron is a very consistent player.

Note, if you look carefully at the steals plot above, you see an extreme outlier at about 1 steal per game. That player is Zach Lavine. If you want to be sure to get a steal per game, you should have Zach Lavine on your team.

Why is Miami the worst team at drives? (March 2017)

Driving to the basket is standard in the NBA. You can make a layup, get fouled, or kick it out for a 3PT assist. I was curious which teams were good or bad at driving to the basket. It turned out, Miami gets the least points per drive of any team in the NBA (see below). To figure out why, first, note that Dragic is the main "driver" of The Heat.

So why is do Dragic drives not lead to points? First, for the high number of drives Dragic s, he hardly gets fouled. He has committed ~700 drives so far this season, but he gets fouled at the rate of someone who only has ~350 drives. Fouls are by far the most efficient way to end a drive.

Additionally, he is not great at making baskets out of drives. He is not as much of an anomaly here, but he is still on the low end.

So what does Dragic do out of drives? Dragic loves to pass out of drives. He really stands out as someone who loves to pass when he drives. Unfortunately, he is not very efficient in turning those passes in to assists. Perhaps he needs to commit to the drive more often.

Which teams run the wrong plays? (February 2017)

Intuition says certain play-types are not as efficient as others (for example, post-ups probably aren't very efficient). Here we break down each play-type in the NBA and analyze each play type's efficiency (points per possession, PPP).

First, How efficient is each play? For this whole analysis, we are using points per possession (PPP) as the measure of efficiency. Below we are plotting each teams PPP for each play type. What we see matches canonical intuition: Cutting to the basket and plays in transition are very efficient. Isolation plays and post-ups are not efficient. Somewhat surprisingly, Pick and Rolls where the ball handler finishes is not an efficient play (regardless of team).

I thought this could be used to identify teams that are running too many inefficient plays or teams that are not running enough efficient plays.

For example, are there teams that run too many isolation plays (a classically inefficient play)? Mavericks, Knicks, and Clippers stand out as teams that over-utilize isolation plays for how effective they are. (I would argue all teams over-utilize isolation plays, but these are the stand-out teams).

Are there teams that under-utilize certain plays? Teams that are very effective at a certain play type, but hardly use it? The Thunder, Timberwolves, Heat, and Jazz are all very effective at plays coming off screens, but they hardly use it. Perhaps they should try to utilize this more often. The Warriors are an example of a team that uses off-screen plays a lot, while being efficient at it.

Below, I'm plotting the distribution of all play-types points per possession and frequency (for each team). This is an overly busy graph, but it is nice to see the distribution of teams. Also, if you want to dig into the actual numbers, the code can be found in analysis/play-type-ppp/playPPP.py

Shot Clock Analysis (January 2017)

Casually watching the NBA, you notice that as the shot clock wears down, teams settle for worse shots. Here we analyze why this happens.

First, do teams actually take worse shots when the shot clock winds down? If we look at eFG% for all shots taken this season binned by time left on the shot clock, we see that teams take much worse shots when the shot clock is near 0. eFG% drops by over 10% late in the shot clock.

Why is this the case? When we break down what shots players are taking, we see that late in the shot clock players take more mid range shots (a low eFG% shot). This is at the expense of shots in the restricted area. Interestingly, 3PT shots are relatively unaffected by the shotclock. To summarize: as the shot clock winds down, players shift from taking shots in the restricted area to the mid range.

I thought it would be interesting to look at teams that are unaffected by the shotclock. What I found was that nearly every team drops in eFG% as the shot clock runs down. However, The Spurs and The Bucks maintain their eFG% as the shot runs down. Below, I am plotting two teams eFG% side by side. You see that Portland is affected by the shot clock (like most teams), while Milwaukee is relatively unaffected.

Westbrook Rebounding (January 2017)

As of today, Russell Westbrook is averaging a triple double:

30.8 PTS
10.5 AST
10.7 REB

How did he add 3 extra rebounds per game to his stat line? He's just soaking up uncontested rebounds.

When we look at uncontested rebounds, Westbrook is a clear outlier. When we look at the rest of OKC, it appears he's taking uncontested rebounds from Adams and Kanter to pad his stats.

Uncontested rebounds is a useless stat that pads your "Rebounds" in the box score. Using the linear relation in the plot above, we can correct each players rebounds stats. For example, Russell Westbrook gets 8.3 uncontested rebounds a game, but given he rebounding ability (contested rebounds per game), we can estimate that he should only get 3.7 uncontested rebounds a game. Using this correction, we see can now look at how many total rebounds players on OKC should get:

Russell Westbrook should get 5.8 rebounds per game (instead of 10.7)
Enis Kanter should get 8.3 rebounds per game (instead of 6.6)
Steven Adams should get 9.4 rebounds per game (instead of 7.6)

Defender Distance (November 2017)

A 3 PT shot is an efficienct shot, but how is it affected by defenders? Here, we look at the affect on 3PT percentage when a defender is nearby, using a dataset containing the the nearest defender (and distance from the shooter) for all shots in the 2014-2015 season.

When we bin all 3PT shots by their nearest defender and plot the 3PT FG%, it is obvious that a defended 3PT shot is a bad shot. If a defender is within 3 FT of the shooter, the 3PT FG% is less than 30%. That means, really, you should only be taking semi-defended and open 3s.

It'll be interesting to see which shooters are unaffected by the presence/absence of close defenders.

Aging Player's Efficiency (April 2017)

What happens to a player's efficiency as they age? Does their offense or defense start falling off first? Here, we combine data from 2000-2016 NBA seasons to address this question.

As players age, their offensive efficiency goes up, while their defensive efficiency goes down. My guess is, players improve on offense due to technical and intellectual gains, but decrease on defense because they exert less effort on the defensive end.

Do players get smarter as they age? Players commit less turnovers and commit less fouls as they age.

How do player's offensive game change? Players take less shots and start assisting more as they age.

Steal Efficiency (May 2017)

How efficient are players at steals? When they attempt a steal, is it successful? Unfortunately, there is no "attempted steal" box score.

The NBA tracks "Deflections" though. A deflection is as close as I can get to an "unsuccessful steal". With that in mind, I plotted each player's steals and deflections from last season.

As expected, if you have more deflections you get more steals. But Kelly Oubre is an extreme outlier. He gets tons of deflections, but didn't get as many steals as you would expect.

Timeout Conversion (May 2017)

The coach's time to shine is during a timeout. I recently heard a media member say something along the lines of "If Brad Steven's got 100 timeouts a game, the celtics would score on 90 of them". Let's test that claim.

Here, I am looking at every possession after a coach calls a timeout. What percentage of the time did they score? The Raptors are the best offensive team coming out of timeouts. The Lakers seem to struggle.

What about drawing up defensive schemes? Below, I am looking at what percentage of time a team can get a defensive stop coming out of a timeout. The Atlanta Hawks are the best defensive team coming out of timeouts.

If anyone knows what goes on during a Lakers timeout, please send me an email. I am very curious.

Rim Protection (January 2017)

The idea for this analysis came through discussions with Justin Palmer

There is lots of talk that Portland Trail Blazers currently need a rim protector. But with Dame and CJ being such poor perimeter defenders (letting everyone past them), could a good rim protector pick up the slack? At the heart of the question, is a good rim protector a product of their team? Or is rim protection an individual skill?

Here, we addressed this question by comparing centers that stayed on the same team for two consecutive years and rim protectors that changed teams. The statistic used to quantify rim protection is opponent FG% at the rim.

What we see is when a center stays on the same team, their rim protecting abilities is consistent from one year to the next (r=0.54). This has been previously documented.

Interestingly though, when a center changes teams, his rim protecting ability is essentially uncorrelated with his previous year's ability (r=0.20). This dichotomy indicates that the team defensive system plays a large part of how well a center can defend the rim.

Perhaps Festus (when healthy) may not be able to solve the Blazers's problems.

Tanking (December 2016)

If an NBA team "tanks" and gets a poor record, they are more likely to get the first pick in the following draft. But how does the first pick change the course of a franchise? Here we analyze the ten year trajectory of all teams since 1990 that obtained the first draft pick.

How does a team's win percentage change after getting the top pick? Teams peak 4 years after getting the first pick, and their regular season win percentage goes up by 30% on average. That is 25 extra regular season wins!

What about making the playoffs? Again, teams peak four years after tanking. The results are striking though. Nearly 70% of teams made the playoffs 4 years after tanking. The percentage of tanking teams making the playoffs stays above 50% for years 4-9 after tanking.

Empirical Bayes estimation of NBA statistics (January 2017)

There is a very common problem after the first month of the NBA season. How do you estimate someone's ability to shoot 3PT shots after they have taken there first 10 3PT attempts? Maybe they only made 1, or maybe they made 9. Either way, it seems unreasonable to use their current 3PT% as your best guess for what their 3PT% will be for the remainder of the season.

Previously, we looked at James-Stein estimation of NBA statistics to solve this problem (see below). Here, we use a method called Empirical Bayes.

So far into the season, how are players distributed when it comes to 3PT%?

In Empirical Bayes, we fit a beta distribution to the empirical distribution and use this as our prior prediction for players probability. What this means is that our starting point for estimating a players 3PT% is to just ask "Well, how do most players in the NBA shoot 3PT shots?"

Using this as our starting point, we use each player's 3 point attempts and 3 point made shots as evidence to update our prior belief. This means if a player has made a ton of 3PT shots, we update our estimate of their 3PT% to reflect that.

In essence, Empirical Bayes accounts for small sample sizes by dialing each players 3PT% towards the league average. Just like James-Stein estimation, the less shots they've taken, the closer to league average they are estimated. If you've made 9 out of 10 shots from 3PT, that will likely not keep up for the rest of the season and your Empirical Bayes 3PT% estimate will reflect that.

As you can see, most players are dialed toward the league average.

Here are some interesting results for players's Empirical Bayes 3PT% estimate:

Even though James Jones has the highest 3PT% (51%), he has only taken 28 3's, so his empirical bayes estimate is only 43%.
George Hill has the highest 3PT% Estimate (44%)
The other players in the top 5 are Kyle Lowry, Danny Green, and JJ Redick

Empirical Bayes can do more than estimate your 3PT%. It can also provide credible intervals (similar to confidence intervals). For example, George Hill's 3PT% estimate is 43%, but his 95% credible interval spans from 36%-52%. As the season goes on and he takes more 3s, the credible interval will get tighter and tighter.

James-Stein estimation of NBA statistics (November 2016)

In fact, commentators frequently say things like, "He's currently shooting 55% from 3, but that should regress to the mean". They don't really know what they are talking about, but there intuition is certainly correct. The idea is that if a shooter begins shooting an extreme percentage, he is more likely to end up about average than maintain the extreme shooting for the rest of the season.

This is what James-Stein estimation tries to accomplish. In frequentist statistics (Maximum likelihood estimation), you would assume every shooter's true 3PT% is whatever he is currently shooting. However, James-Stein estimation says, actually, we need to dial it towards the league average. Importantly, the extent to which we dial it back to the league average depends on how far from the mean the player is shooting and also how many shots the player has taken.

So if a player is shooting 35% from 3 after taking 300 shots, James-Stein estimation will say he's probably about a 35% 3PT shooter. On the other hand if a player is shooting 80% from 3, but has only taken 5 shots, James-Stein estimation will more say that players true 3PT% is much less than 80%.

In PER/PER.py we calculate the James-Stein estimated 3PT% for each player in the NBA using their 3PT shot data so far this season. Details of the calculation can be found in:

Efron, Bradley, and Carl N. Morris. Stein's paradox in statistics. WH Freeman, 1977.

James-stein estimation is given by:

where y is a vector of player's current 3PT% and the number of players m. Notice how y is shrunk towards the origin.

Here we have plotted the players empirical 3PT% (blue) and James-Stein 3PT% estimation (green). As you can see, the extremes are dialed in towards the league average.

It also helps to look at individual cases. J.J. Redick is a great 3PT shooter. So far he is shooting 49%, making 34/69 shots. Since he has taken a healthy amount of shots, James-Stein estimation only dials him back to 43%. On the other hand, Serge Ibaka is also making nearly 49%, but only on 38 attempts. James-Stein dials his 3PT% estimation all the way down to 37%.

Here are some interesting estimates of 3PT%:

Player Efficiency Rating (PER) (November 2016)

Player efficiency rating (PER) is a linear combination of box score statistics which is commonly used as a single number to describe how valuable a player is. Here, we look at a number of properties of PER (from the 2015-16 season).

PER is normalized so that the league average for a given season is 15. But what is the actual distribution of PER?

What we see is that the distribution of PER isn't actually centered around 15. The fact that the mean is 15 is due to players with extremely high PERs bring up the average. Instead, the bulk of players are centered around 12-14. In fact, the distribution is quite tight, with a standard deviation of ~4.

With PER, we can see when players peak during their career:

For the high-end talent, players peak in their late 20's and begin a fairly linear decline.

One common criticism of PER is that Offense is heavily favored. When we break down by position, we see that Point Gaurds and Centers are favorite. This is due to PER weighting things like assists and rebounds heavily, which are biased towards those positions.

Lastly, one might expect to use players with higher PERs more than players with low PERs (not a very complicated hypothesis). Indeed, this seems to be true, with a linear correlation between usage and PER.

What is interesting though, is which players are underused or overused according to their PER.

DeMarcus Cousins, while being a great player, was far over-used according to his PER. This is likely due to him being on a dysfunctional team. Kobe was obviously the most used player in the NBA (by this metric, but also by the eye test). Everything was running though him and he was not very good.

Alternatively, Boban and Whiteside were underused. I attribute this to their inexperience and coaches not yet trusting them enough to run plays for.