Returning a kick or punt for a touchdown has never been rarer in the NFL with one coming on average every 30 games.

So when there’s a moment like there was Sunday when Nyheim Hines electrified the crowd in Buffalo with not one but two kickoff return scores in the same game, figuring out how it happened becomes a key question.

The NFL NextGen Stats group has teamed up with machine learning engineers from Amazon Web Services to develop the first advanced stats model focused on kick and punt returns to quantify it in a new statistic set to be officially released to the public Thursday.

Building on the work the NFL has done with AWS with player-tracking data to quantify statistics like completion percentage over expectation, rushing yards over expectation, yards after catch over expectation and defensive coverage classification, it now is doing the same on returns.

“When we talk about what’s missing from the special teams game, it’s really all of the other pieces of special teams that impact the field position game, which more or less impacts the strategy of the game tremendously,” said Mike Band, the senior manager for NFL Next Gen Stats Research & Analytics. “We feel that expected return yards is one of those areas that can shore up that missing piece.”

By tracking the location, speed and acceleration of all the players at the moment a returner fields the ball through player tracking data gathered by chips in the players’ pads and comparing it to historical data from the previous four seasons, the model is then able to estimate the probability distribution of how many yards a returner could be expected to gain.

From there, the NextGen Stats group can measure stats like how many yards over expectation a player gets on each return, how often they are successful, how many expected yards they gave up on fair catches and decision-making on touchbacks and turn it into a ranking score for both punt and kick returners.

“We’re trying to put it in the terms that that would be as digestible as possible in the short amount of time,” Band said. “Our stuff goes on screens, so we need to keep our stuff concise as possible. So we’ll call this a punt return score on a per play basis.”

Taking Hines’ opening kickoff return last week for the Bills as an example, the model estimated that he was expected to get a 21-yard return when he caught the ball at the 4. Instead, he gained 75 yards over expectation. The model gave him a 0.6% chance of scoring on that return, making it the third most improbable kick return TD since 2018.

“That’s the kind of stuff we can do,” said Vasi Philomin, AWS vice president of Machine Learning and AI Services. “It’s all based on the fan experience and enhancing it, looking at what would make the experience even better for them. We look at the demand for statistics that they have and be able to present something that’s very meaningful.”

Hines’ second return TD that made him the first player since 2010 with two in one game, was less improbable with an estimated 1.1% chance of scoring when he fielded it 1 yard deep in the end zone. The return was expected to gain 25 yards.

Those two big returns helped boost Hines’ ranking on the season. His 138 yards over expectation ranked second best to Keisean Nixon’s 214 for Green Bay. But Hines’ average of 7.2 yards over expectation per kick return ranked No. 1, ahead of Seattle’s Godwin Igwebuike (plus-7) and Nixon (plus-6.1).

On punts, Patriots rookie Marcus Jones was tops in the league with 119 more yards than expected, thanks in part to his 84-yard, game-winning return against the Jets in Week 11 that had a 0.5% chance of scoring and generated 72 yards over expectation.

While Jones was more boom or bust, the Chargers’ DeAndre Carter was the most consistent punt returner with a 69% success rate on his returns.