"What AI systems have discovered is that focusing on late-round games is more effective than starting with the Round of 64 games and moving forward," Sheldon H. Jacobson writes. (Gene J. Puskar/The Associated Press)

Opinion editor’s note: Strib Voices publishes a mix of guest commentaries online and in print each day. To contribute, click here.

•••

The madness of March is here. College basketball teams from across the nation are preparing to battle it out on the courts, to soundly beat their competition or pull unfathomable upsets. The mantra of “survive and advance” in the nation’s most watched and followed single elimination college basketball tournament is what every coach has on their mind.

Fans have organized office pools to see whose bracket picks the most upsets or the most overall winners, with some keeping the focus on the Round of 64 games this Thursday and Friday. “March Madness flu” used to be why people called in sick on those days. Yet with an app available to cover every game, the biggest casualty of the tournament on these days is lost productivity in the workplace.

Artificial Intelligence systems use data, something that the 39 years of tournaments since 1985 offer an abundance of. This is when the tournament went to its 63-game format. With all such data, shouldn’t an AI system be able to pick the winner of every game?

The simple answer is, no. Every game is a complex combination of possessions and player interactions, with each such interaction carrying uncertainty. This means that every top ranked team, the so-called Goliaths seeded No. 1 or No. 2, can be upset by the Davids in the field, those seeded No. 15 or No. 16.

Just ask the University of Virginia in 2018, when they lost as a No. 1 seed to No. 16 University of Maryland Baltimore County, or Purdue in 2023, when they lost as a No. 1 seed to No. 16 Fairleigh Dickinson University.

No. 15 seeds have enjoyed more success, with 11 wins in 156 games since 1985. However, seven of those 11 upsets have occurred since 2012. No AI system can reliably find such “needles in haystacks.”

So what can AI systems do to help fans put together sensible brackets? There are several scoring systems used to evaluate brackets, with all rewarding brackets that get the later round games correct. ESPN, Yahoo, CBS Sports, FoxSports and NCAA.com all share a common scoring system, doubling the points earned per game with each subsequent round.

AI systems use such information and focus their attention on the latter rounds to sample from all possible game outcomes. However, with over 9.2 quintillion possible brackets, even AI systems cannot sort through them all.

But they do not need to. Using historical performance data, which contains information on how seeds have advanced in prior tournaments, the odds of selecting all 63 games correctly are around 8.6 billion-to-1, still quite large, though not as daunting as 9.2 quintillion-to-1.

What AI systems have discovered is that focusing on late-round games is more effective than starting with the Round of 64 games and moving forward. But which late round? The ideal rounds are the Elite Eight or the Final Four. Picking these games first, and then filling in your brackets around these picks give you the highest likelihood of finding high scoring brackets amongst a pool of brackets generated.

One benefit of the AI system strategy is that it reduces the likelihood of picking too many early round upsets, or if they do get picked, they are eliminated early enough to not degrade the entire bracket.

How well does such an AI system strategy work in practice? We implemented it in bracket simulators for the 2021 through 2024 tournaments, generating one million brackets each year prior to the start of the tournaments. We then looked at the highest scoring brackets and compared them to the highest scoring brackets in the ESPN Tournament Challenge Leaderboards.

The best AI system strategy brackets topped or matched the ESPN winners in 2021 and 2024, and were in the top 100 in 2022 and 2023.

What the AI system strategy cannot do is reliably pick a single bracket that would win the ESPN Tournament Challenge or any bracket pool. If it had that ability, everyone would have the same bracket, and the games being played would be superfluous. Of course, games are played on the court, not on a computer, so there are many brackets that an AI system strategy would consider reasonable based on the historical data. What it cannot not do is specify one such bracket as the outcome of the tournament.

However, given enough brackets, an AI system can create a set of brackets whereby one of the brackets will score well with a reasonable probability.

The takeaway is that AI is best used to identify strategies that can lead to sensible brackets. Yet even sensible brackets get busted. That is what makes March Madness exciting, even for those who do not follow college basketball. When teams with nicknames like Bulldogs, Ducks and Wildcats battle it out on the court, there is something for everyone to enjoy. Let the games begin, with AI guiding your every bracket!

Sheldon H. Jacobson, Ph.D., is a professor of computer science in the Grainger College of Engineering at the University of Illinois Urbana-Champaign. He applies his expertise in data-driven risk-based decision-making to evaluate and inform public policy. He is the founder of Bracketodds, a STEM learning lab showcasing the mathematics of March Madness.