Location Analytics Applied to the NFL

Location data generated from cell phones powers many of the ad-tech products with which we all work every day, from in-person attribution to targeting segments based on visitation patterns. Earlier this year, I got to work with data generated from a completely different source — professional football players.

The NFL, following in the footsteps of the MLB and NBA, now collects precise location data for every player on every play. Some basic stats based on this data — like top speed and distance travelled — have already begun appearing on TV broadcasts. Descriptive stats like those are a good starting point, but teams are interested in unearthing deeper insights to gain a competitive advantage.

To highlight the types of analysis that are possible with location data, the NFL organized the Big Data Bowl, an analytics contest that asked college students and analytics professionals to use the NFL’s player location data to surface insights. When I heard about the contest, I saw it as a chance to combine the type of techniques I use every day with my favorite sport, football.

In my submission to the Big Data Bowl, I analyzed the effectiveness of route combinations. In other words, I used the location of the players to determine the optimal target on each play.

The first step was to identify the routes on each play. In the NFL, this task is painstakingly done by hand by “Quality Control” coaches, who work 16-hour days during the season to help their teams prepare for the next game. I needed to identify 35,000 routes and only had a handful of nights and weekends to do it, so that type of manual review would not work. Instead, leveraging my experience working with location trace data from cell phones, I trained a neural network to recognize routes. After all, identifying routes is similar to identifying visits; in both cases, you need to build features from the raw location data and then train a system to separate the signal from the noise. When building my route recognition network (“RouteNet”, for short), I translated the raw location trace data into images and used convolutional layers to pick up on the characteristic shape of each type of route.

This plot shows all wide receiver routes run during week five of the 2017 NFL Season. The red lines are those classified by RouteNet as matching the label at the top of the image. As you can see, RouteNet was able to accurately classify routes.

After assigning route labels, I used the location of the players to create route combinations and then computed each combination’s effectiveness. To put it in terms more familiar to ad tech insiders, I identified the tactics that best optimized conversions. One finding was that “Corner’” routes outperformed “Go” routes, both when run by themselves and when paired with other routes. Another was that Steve Spurrier’s favorite play from his days coaching the Florida Gators is still effective in the NFL today—a surprising insight because football has changed a lot in the past 25 years.

After submitting my paper, I was named one of the four open-division finalists and flown out to Indianapolis to present my findings at the NFL Combine. About 150 people attended the presentation, including reps from all but a handful of NFL teams. After each finalist presented, I was named the overall winner.

The challenge of extracting insights from location data is not unique to NFL teams. I have been working on that problem for years at Placed, and it was fun to learn that the same kind of tools we use every day to analyze location data can help NFL teams win on Sunday.

Nate has been working with data for more than 10 years. At Placed, he has used raw location data to measure the effectiveness of advertising campaigns and to create targeting segments. He’s thankful to the NFL for helping him justify all the time he spends watching Michigan football games. 

Tags:
Next Post

No Matter What Business You’re In, You Are In The Business of Selling