Few things are more controversial in sports than “momentum.”
Most observers insist it’s real, a concept that’s intangible but nevertheless observed and felt as sporting events transpire. The hot-shooting basketball player keeps making shots. The player who’s 4-for-4 is more likely to get that big hit in his fifth plate appearance. The team “on its heels” is likely to fall apart. Announcers constantly call attention to the seizing of momentum.
But is it something our brains just invented? Numerous studies and articles have considered the question, many arriving at the same answer: It doesn’t exist, no matter how much we believe otherwise.
“Psychologists will tell you there’s an evolutionary advantage to recognizing a pattern even when it doesn’t exist,” said Paul Roebber, who teaches data science classes and specializes in weather patterns at the University of Wisconsin-Milwaukee.
“Sometimes these are false patterns, and the classic example they talk about is if you’re at the edge of savannah and see the grass is waving, it’s 99% likely to be the wind, but if it’s a lion stalking you, it’s better to assume it’s a lion stalking you because that will be the end of you if it is. We are programmed to recognize patterns even if they don’t exist.”
Can momentum be identified as tangible in football?
But here’s the thing: Roebber has done programming to prove the opposite.
Roebber, graduate student Bryan M. Burlingame and Roebber’s football-minded friend Anthony deWinter set out to discover if the intangible could be identified as tangible. Their findings were published in PLOS ONE research publication and touted by UWM with a fun pronouncement: “The fans are right.”
“There had been some studies that tried to verify if ‘the hot hand’ exists or refute how it exists, and I wasn’t satisfied with the results because I felt like they weren’t really defining it correctly,” Roebber said. “I was thinking about the team’s performance and if the team gained momentum, not the individual player.”
Roebber, a Boston native who’s been in Milwaukee since 1994, first started working on the project when Burlingame, a football fan, suggested they explore it as a side project to something else.
“He was interested in football, but unfortunately he’s a Lions fan,” Roebber joked (this was perhaps funnier before the Packers lost to the Lions this year … sometimes it isn’t just the wind moving the grass of the savannah, after all).
What the momentum project required
The project required two first steps: Developing a “win probability” calculator that recognized when a team’s chances to win a game improved, and defining momentum.
Once they figured out the former, they created a neural network and fed it NFL play-by-play data from 2002-12. They defined momentum as seeing win probability increase over the span of three successive series (offensive or defensive). If the win probability had steadily improved, a team had achieved positive momentum; if probability had steadily decreased, that was negative momentum.
They used seven different inputs to arrive at their win probability. The most important elements to consider, Roebber said, were the score in the game at any point in time, the point spread (as in, the expectation one team is better than the other, thus lending insight into a team’s likelihood of winning) and where the team was on the field.
Roebber said the benefit of “momentum” moments on a team’s chances of winning was not just statistically significant, but overwhelmingly so.
“Teams that eventually win have about 14% more of these momentum developments or these streaks in these game than teams that don’t win, which certainly seems to have a relevance as to whether you win the game,” Roebber said.
That might seem patently obvious — a team with three straight good series has a better chance of winning. But remember, win probability is relative to where a team stands in the game — teams dominating a game won’t necessarily achieve Roebber’s definition of “momentum” because the bar to improve their likelihood of winning is already high.
Roebber said his software can project if a team is going to win based on the elements of the game. The projections are roughly 80% accurate after the first quarter, then increase to 90% by the fourth.
“It’s not perfect, of course,” Roebber said. “The famous Super Bowl game between the Atlanta Falcons and New England Patriots, late in the third quarter Atlanta had a 99% win probability … I call those ‘black swan events’; you can’t predict those sorts of things.”
Momentum is ‘function of all the players,’ analysis says
Can he go one step further, and predict when momentum events are likely to occur? He’s tried that, too.
“We built another neural network where we used win probability as input … we found it does a pretty good job finding out whether a streak is going to happen before it happens,” he said. “There’s more uncertainty (than with the other predictive model), and it seems like the most important thing for that to happen is if you’re trailing, but only slightly, have a lot of time on the clock and less total scoring (in the game).”
He’s considered looking at other sports, but the NFL seems more conducive to overall evaluation and not player-specific evaluation.
“When you’re talking about a team sport like football, my view is that momentum is a function of all the players,” he said to the UWM Report. “And so, you really need to look at the collective performance of the team.”