Simulating The Madness of March College Basketball: A Decade Later, with AI at Our Side
The NCAA College Basketball Tournament and Data Analytics: A Fun Exercise in Predicting the Unexpected
Every March, the NCAA Basketball Tournament seemingly becomes a National ritual. Even people who haven't watched a single college basketball game all season suddenly find themselves filling out brackets—sometimes choosing winners based on preferences in school mascots, jersey colors, or gut feelings. It’s a cultural moment where data analysts, die-hard fans, and complete novices all try to make sense of the same chaos.
Back in 2015, we decided to get in on the fun—not to win a pool or showcase basketball expertise, but to try a structured, data-driven approach to trying to predict outcomes in a famously unpredictable event. We built a relatively simple simulation that used a performance metric called RPI (Ratings Percentage Index) and factored in historical patterns around how many upsets typically occur in each round. (For those unfamiliar, upsets—when lower-seeded teams beat higher-seeded ones—are a defining feature of the tournament and a big part of what makes it so fun.)
Ten years later, we revisited the challenge—this time with AI in the mix. The goal wasn’t perfection, but progress: Could we use AI to help us quickly build a more nuanced, dynamic, and realistic simulation model?
What’s New in Our 2025 Simulation?
Our 2025 version builds on the same foundational idea from 2015—use a performance metric, consider historical data on how many upsets tend to happen per round, and run a full tournament simulation. But this time, we upgraded the data and introduced more nuance and unpredictability—thanks in part to the flexibility and speed AI gave us during development.
✅ A Smarter Performance Metric
We replaced RPI with BPI (Basketball Power Index), a more advanced and predictive performance metric developed by ESPN. Unlike RPI, which was based primarily on win/loss records and strength of schedule, BPI incorporates efficiency data and scoring margin, making it better suited for forecasting matchups.
✅ Historical Upset Ranges with Randomization
We used historical data to guide the number of upsets per round, but we didn’t fix those numbers in the simulation algorithm. Instead, we randomized within historical ranges of upsets per round. Some tournaments are full of surprises, others are more chalk-heavy (where higher seeds dominate). Our model now reflects that variability.
✅ Introducing “Shocker” Upsets
The NCAA Tournament almost always delivers a few jaw-dropping early-round results. To simulate that reality, we introduced “shocker” upsets—results that ignore the numbers entirely. These rare surprises represent cold shooting streaks, hot hands, missed free throws, momentum swings, and all the other intangible chaos that defines March Madness.
✅ Late-Round Chaos (within reason)
We also allowed for some variability deeper into the tournament. While late-round upsets are less common, it’s not unheard of for a lower-seeded team to make a deep run. Our model builds in some (rare) chances for those Cinderellas to keep dancing.
AI as a Collaborative Partner
AI didn’t generate our simulation on its own, and in a few weeks time we certainly won't find that it 'solved’ March Madness (the odds of filling out a perfect bracket are, at best, 1 in 120.2 billion (read this interesting NCAA piece on The Absurd Odds of A Perfect Bracket)
. Instead, it served as a collaborative tool, accelerating the development of a more sophisticated model by:
- Quickly exploring new logic structures.
- Iterating through variations with enhanced creativity.
- Identifying and addressing edge cases more efficiently.
However, it's crucial to note that AI functioned best as a collaborator, not a replacement. Contrary to popular narratives suggesting that AI is poised to take all of our jobs, our experience indicates that, today, its greatest value seems to be in augmenting the capabilities of experienced professionals. This synergy enables faster, more creative, and more flexible problem-solving— whether you're modeling a basketball tournament or helping businesses interpret feedback, identify patterns, and stay ahead of shifting guest/customer expectations.
Let the Games Begin
So... our v2025 GuestInsight AI March Madness Simulator has generated our bracket and we are eager to see how the real tournament unfolds. If you're interested in following along, we've published our live bracket here so you can track how our simulation performs.
If you’re entered in a March Madness bracket pool yourself, good luck! And, let us know after the tournament ends if you beat us.