The day before the NFL announced its winners of its annual Big Data Bowl competition, Marc Richards went to his Twitter page and set up a notification for any new tweets from league analytics director Michael Lopez, based purely on hope and a hunch.
The Plymouth native sent a note to Jack Werner, his friend from St. Olaf and one of three Minnesotans on his Big Data Bowl team, saying, "I don't know — we might have a chance here."
Minutes after Lopez announced the five winners the morning of Feb. 5, Richards relayed the news and sent the group into a tizzy.
"We won!" he wrote, in a text message linking to Lopez's announcement.
"OMG," Werner responded.
Sam Walczak, the third Minnesotan on the team, responded, "Hold on … so do we win $15K?"
What began as a labor of love for the three St. Olaf friends and Wei Peng — whom Richards met in their Ph.D. statistics program at the University of Pittsburgh — turned into a chance to showcase their work nationally. The NFL named the group one of its five finalists in the open division of the third Big Data Bowl, a competition in advanced sports analytics.
The group did win $15,000 and will present its work, which used NFL player tracking data to identify types of defensive coverages and evaluate player performance, to league executives March 18 with a chance to win another $10,000. The presentation will be streamed on YouTube, Twitch and the NFL app.
Richards, Walczak and Werner attended high school in the Twin Cities suburbs and met as undergraduates at St. Olaf, where they were part of a group that bonded over an interest in sports analytics. They started Model 284, a website to showcase some of their data science work on pro sports.
Peng, who grew up in China, was a soccer fan and hadn't watched much football until Richards invited him to a party for the Chiefs-49ers Super Bowl in 2020.
"He likes to say that I tricked him into watching the NFL and football, so we could do the Big Data Bowl competition together," Richards said. "We just were like, 'With COVID times, we need something to do rather than going out to eat on the weekends. Let's try to tackle this.' "
• • •
The NFL gave Big Data Bowl contestants access to player tracking data from the 2018 season and asked them to develop ways to measure defensive performance.
Peng, Richards, Walczak and Werner first built a model to determine if a player was in man or zone coverage, based on how they were lined up at the snap and how they moved once the ball was snapped. That allowed them to group coverage assignments into different clusters, build a model predicting how many passes should be completed based on how close a defender is in coverage and evaluate player performance against those expectations.
They met weekly over Zoom to discuss their project and divvy up individual assignments. Peng developed an algorithm that would help the group identify defensive assignments at each frame of a given play.
The pandemic gave Peng and Richards an extended winter break to put extra hours into the project, while Walczak and Werner contributed several hours a week around their jobs for Holborn Corp. and Wells Fargo in the Twin Cities.
How many total hours did the group spend on it from October to their final submission in January?
"Enough to be less than minimum wage when you break it down," Richards said with a laugh.
Money, though, wasn't the primary motivator.
"For me, it was a nice way to keep in touch with friends who are now in different cities, and who you're not really seeing anyway because of the pandemic," Werner said. "It's nice to have a weekly project to work on."
After he announced the winners on Twitter, Lopez pointed out a diagram of a Matt Ryan-to-Julio Jones completion from the group's project, which predicted where Ryan would go with the ball based on where the Atlanta Falcons quarterback was looking during each tenth of a second in the play.
Lopez tweeted: "How cool is this finding? You can see target probability for the Falcons change ** because of Matt Ryan's orientation **. There are about 8 or 10 awesome findings from this year's projects that I want to share, but this one had me do a double take."
During the group's final presentation to NFL staffers next month, it figures to face a question coaches often raise about data-based models for evaluating players: How can anybody say whether a player was right or wrong without knowing the player's assignment on a given play?
"We're not going to be 100 percent, like, 'This is exactly what this defender is supposed to do.' But what we're doing here is, we're creating models that now somebody doesn't have to sit and watch all the film, or we can get rid of some of the work to do that," Richards said. "Maybe we're not 100 percent right, but we're like 90 percent or 85. We're better than 50-50, we're getting closer to the right answer and we're compensating time to do that."
• • •
After Peng graduates this spring, he is joining a Pittsburgh startup working with data and machine learning to detect cancer in its early stages. His football fandom is still relatively new, but after the Big Data Bowl, he watched this year's Super Bowl at a more nuanced level.
"Now, when I watch a game, I'm trying to see whether we have the defensive assignments [correct] in our project," Peng said. "I want to see whether this assignment makes sense or not in practice. I'm more focused on the defensive team instead of the offense."
Richards, who had previously participated in an NHL data competition, will spend the summer as a data science intern for the Oklahoma City Thunder, and Walczak is working on a master's degree in data science at the University of Minnesota.
It remains to be seen whether anyone in the group will pursue a career in sports analytics, but the Big Data Bowl probably won't be the last time any of them use their skills to provide a new perspective on the games they watch.
"Just like the Big Data Bowl stuff, I enjoy developing new algorithms, or just taking data and applying algorithms to solve real-world problems," Richards said.