Go for It? New Decision Data Makes NFL Next Gen Stats a Growing Part of the Live Game Broadcast

AWS Machine Learning Solutions Lab crafts new analytics to offer storylines within the game

Perhaps the highlight of a raucous Week 2 in the NFL came late on Sunday evening when Baltimore Ravens quarterback Lamar Jackson lunged across the marker on a fourth-and-1 to seal a statement win over the defending AFC Champion Kansas City Chiefs.

It was an aggressive call — one viewed by many, though, as a bit of a no-brainer. The latest data from NFL Next Gen Stats would agree.

According to Next Gen Stats, a program developed internally by the league in 2016 with tracking-tech-provider Zebra Technologies and running on Amazon Web Services (AWS) infrastructure, going for it in that spot of the game was the preferred call for the Ravens by a difference of 24% in win-probability value vs. a punt. In addition, the numbers gave Baltimore a 75% chance of converting the yardage necessary to pick up the first down in that specific play.

This kind of fourth-down–conversion data, along with similar metrics around whether it’s the right time to go for a two-point conversion, is a key element of the latest offering from the NFL Next Gen Stats team: the Next Gen Stats Decision Guide powered by AWS.

“Next Gen Stats has come a long way,” says Josh Helmrich, director, strategy and business development, NFL. “Machine learning has allowed us to take statistics to the next level. When we were first getting started [in 2018], we were measuring things like yards covered, speed, and who is on the field. Some of those things, we believe, still have merit and add value, but one of the key tenets of Next Gen [Stats] has always been that this technology can help everyone. We want to develop many statistics and get as much good information out of the game to benefit coaches, players, fans — everyone top to bottom. The deeper that tool kit becomes, the more people you can help.”

Data of this nature is already proving valuable to broadcasters. The data in the Next Gen Stats Decision Guide allows them to use hard probabilities rooted in historical data to give context to some of the most debated moments in the game before they happen.

According to Helmrich, the partnership with AWS has been the catalyst for taking the NFL Next Gen Stats concept to the next level. The work of the AWS Machine Learning Solutions Lab led by Priya Ponnapalli, senior manager, applied science, AWS, resulted in development of the Next Gen Stats Decision Guide, which is built on a series of machine-learning models using Amazon SageMaker.

According to the league, the decision equation focuses on two main components: win probability, which attempts to quantify how much the game’s final outcome will change in the event of each play, and conversion probability, which quantifies the likelihood of the offense’s converting on the two-point conversion or fourth-down play at hand.


“You need context to make these statistics impactful,” says Helmrich. “Over time, we’ve been able to grow, and, before AWS, we didn’t have all of the machine learning that we do now. We’ve been able to produce more computer-assisted statistics. Obviously, teams can have success or failure in either scenario, but, based on the numbers, what’s the call that gives you the best chance to win?”

It’s a key step for the Next Gen Stats project as a whole. It converts complex algorithms and AI-generated data into concepts that are easy to consume during the game.

Even when the NFL is working directly with broadcast partners CBS, Fox, NBC, ESPN and NFL Network, storytelling is critical for the league in relaying the data’s value. Simply handing over a new trove of raw data isn’t enough. Helmrich says it has become increasingly critical to frame data in a story and craft it into context so that behind-the-scenes personnel can see the value in and feel empowered to produce content for the broadcast in the form of graphics or replay packages. It also helps with buy-in among the on-air talent, who can use the data in their analysis of the game.

“I think the team here has done a really good job of building good content to help get the media folks to buy in,” says Helmrich. “If [the data] isn’t interesting, broadcasters won’t use it. If it’s not reliable and high quality, the teams won’t use it.”

With more data acquired and automated through machine learning, NFL Head of Product and Technology Matt Swensson and a research team including Next Gen Stats analysts Mike Band and Keegan Abdoo and data scientist Adam White have been able to improve an internal tool that makes it easier to relay valuable statistics in real time to broadcasters during the game and makes the data more accessible for broadcasters to mine on their own.

“We need to help them understand what’s new and how it can be best utilized by them on-air,” says Helmrich, noting that researchers from his group are in communication with a graphics operator or a production assistant in the production truck during the game via email or communication applications like Slack. “We help with that a lot. Our research team helps provide specific interesting stories going into a game. Then, in the game, we’ll have a dialogue back and forth with the broadcast partners. They will ask us specific questions and for data that helps explain certain players [or scenarios].”

NFL Next Gen Stats are built through the use of a tracking system installed at all NFL venues. The complete setup includes 20-30 ultra-wideband receivers as well as RFID (radio-frequency identification) tags installed in players’ shoulder pads, on officials, and in balls, pylons, and first-down chains. The league estimates that about 250 sensors are active in a stadium during a game, pumping tracking data to the league’s home office in New York City.

In addition to the Next Gen Stats Decision Guide, the NFL and AWS are also debuting a suite of new stats for the season: Quarterback Expected Rushing Yards, Quarterback Dropback Type, Next Gen Stats Big Play Score, and Expected Fantasy Points.

Password must contain the following:

A lowercase letter

A capital (uppercase) letter

A number

Minimum 8 characters