In our last post we saw that we can use a simple nonparametric statistical method, called the kernel density estimator, to build nice looking scoring charts. As we saw, these charts give a better idea of the distribution of the scorer when compared to the basic shooting chart used for several decades past. In this post, we take a look at how we can turn these spatial charts into spatio-temporal charts; making a richer dataset for us to use.
Spatio-temporal data is a collection of spatial points, like shot locations on a basketball court, and indexes these points by time. The most common usage of spatio-temporal statistics is found in meteorology; when scientists attempt to determine the weather at a given location and future time, given the current weather at every location at the current and previous time points. We can do the same thing for Basketball.
When we consider data in this format of space plus time, we have to primarily do two things: identify a method for calculating meaningful statistics and identify a method for displaying the data and results. Here, we focus on displaying the data, called visualization. The most common method of visualization of spatio-temporal data is the animation.
1. Breaking Down Plays In The NBA
One example of spatio-temporal data is the location of players on a court at a given time. Companies such as SportsVU have cornered the market with several NBA teams, where they install a camera system that tracks objects on the basketball court, sampling at rates of roughly one snapshot per quarter second. Here an object may be the basketball players, the basketball, or the referees. The data is then made available to the team that purchases SportVU’s system and is ready for analysis. Data like this is invaluable to a team that is trying to identify player tendencies that scouts just simply cannot calculate.
For example, let us consider an example from Game One of the NBA Finals in 2014. Here we focus on a series of three plays for the San Antonio Spurs offense as they are defended by the Miami Heat. In each play, the ten player locations and the location of the basketball are extracted from the video. We treat this as eleven correlated spatio-temporal processes during the executed plays. In the attached video, we see a preview of the play and then the extracted coordinates overlaid on the basketball court. The coordinates are read from a file that merely contains the eleven processes, times for each process and the locations at each given time.
In each of the three plays, we see similar style of plays: high screen on guards, rolls from point to post, wings spread to corners. The San Antonio Spurs became notorious for their style of play to be either lay-ups or corner three’s over the past decade. The distribution of scoring for the San Antonio Spurs in our previous post is evidence of that. In each play, the Spurs either score or are fouled at the rim by forcing two particular Miami Heat defenders to make a decision to guard between one of two Spurs players.
A trained scout can easily pick up this information by watching the raw video, however the spatio-temporal data gives us locations to identify how quick those players react. The locations, combined with the times of those locations, allows us to calculate velocities of player rotations, separation distances between players, and rates of close-outs. We can use these values to identify which players are most responsible for the basket to be made.
When calculating these values, we see that in the first clip that Chris Bosh hangs tight on the weak side block the entire duration of the play. When the defense breaks down, Bosh holds on the post player in the short corner, forcing LeBron James to come from the weak side wing to attempt a feeble stop of Tony Parker and Rashard Lewis to cover the middle of the key. The ability for Thiago Splitter to get fouled at the rim was due to Bosh’s inability to rotate over on the play; creating enough space for an easy dump pass for a lay-up.
Similarly, in the second play, Norris Cole sits in “No Man’s Land,” staying out of illegal defense territory, as the play evolves away from his half of the court. Due to the high screen and roll, Cole never manages to cross back into the lane for help defense and an easy lay-up is had by Splitter.
In the third play, Norris Cole finds himself out of position again as this time he jumps a double team and leaves Patty Mills wide open for the basket. Calculating the distances relative to Manu Ginobili’s location of the pass, LeBron James could have stepped out to cover. However this would have forced Cole to cover Splitter in the post; being over 11 feet away.
As we can see, spatio-temporal sets can help identify player tendencies over a period of time. This is just a small example of how to effectively use the datasets.
2. Identifying Shooter Tendencies Over Time: A Curious Case of Luol Deng
One interesting way to look at player shot charts is to look at them over time. For example, we take a look at Luol Deng’s scoring history. Luol Deng was drafted out of Duke University after one year of collegiate action by the Phoenix Suns. Playing no games with the Suns, Deng was traded to the Chicago Bulls and completed his rookie season by starting 45 games, playing in 61 games, and scoring an average of 11.7 points per game.
What’s noticeable about Deng is that he was noted for being an attack the rim kind of player. In his rookie season, Deng made 280 field goals; 31 of which were from behind the arc (11.1% of total field goals). The shooting distribution for Deng was primarily composed of dunks and lay-ups. There barely seems to be a hot spot anywhere for Deng outside of the charge circle.
This type of trend continues for Deng from season to season until the 2010-2011 NBA season. In the 2010-2011 season, Deng entire changes his shooting distribution and integrates three point shooting into the mix.
The change is obvious: The Chicago Bulls changed coaches and philosophy in 2010. After Vinny Del Negro had his coaching tenure ended, the Bulls brought in Tom Thibodeau who allowed Deng to open up his game. Between the Scott Skiles, Jim Boylan, and Del Negro eras we see Deng make between 1 to 32 total three point attempts; taking no more than 117 attempts in his rookie season. Since Thibodeau’s first season as coach, Deng attempted 189 or more per season and makes no less than 57 any season.
As Deng moved to Miami, we see that Deng finds his scoring distribution to match the mold of the Miami Heat offense: lay-ups and corner three’s. All of these changes are easily captured in the animation; displaying the changes in Deng’s scoring opportunities from season to season.
We do not have to restrict ourselves to scoring capabilities of a particular player. Just as in our previous post, we can look at the scoring distributions of teams. Similarly, we can look at the distributions of opponents scoring against a particular team, as well as the distribution of scores for opponents when a particular player is in the game. Grantland has a great post on this from the 2013 – 2014 NBA Season.
Hopefully this gives good reference on relatively simple things we can do with spatio-temporal statistics. This is just the tip of the iceberg when considering sports analytics and measuring player abilities on the court.
2 thoughts on “Spatio-Temporal Data In the NBA”
Pingback: NBA Data Science: Breaking Down NBA Data | Squared Statistics: Analyzing Crime, Sports, and People
Can you start providing some of your code? quality work otherwise