It’s a brave new world. After playing March Madness pools for over forty years and even winning a few (how about NC State in 1984?), I worked with ChatGPT to help me fill out my brackets. Like most people, I watch a few college basketball games – but it’s usually my favorite three or four teams and whoever they’re playing against. I’ll be watching some of the big national games this weekend, but that’s about it.
So when I stare at the yawning line, I usually wonder: Where is High Point University and who are they playing? What about the dreaded 12/5 surprise? Is picking the No. 1 seed too tricky to win? Lots of fear and little insight.
Can a machine outwit forty years of instinct?
I asked how it would create a model to predict the winner of the NCAA tournament. The reaction alone made me say “Wow.”
Here’s the answer: “I can definitely help build one with you: from a simple seed-and-efficiency model to a more serious probability model that uses KenPom-style efficiency, pace, injuries, travel, coaching, and surprise history. I can also help with: Bracket win probabilities, surprise picks by round, Monte Carlo tournament simulation, Elo or logistic regression models, and table-based prediction tools.”
ChatGPT considered team strength on a neutral site, including offensive and defensive efficiency, strength of schedule, consistency and roster strength, among other things. This ultimately led to a series of predictions based on this equation:
These are the considered matches between two teams based on these team ratings and using a logistic model.
Ultimately, the entire tournament was simulated using this technique no more than 50,000 times.
Do you trust the algorithm – or your gut feeling?
I don’t know about you, but this works at a level of processing power and processing power far beyond my capabilities. Then it came back with a brilliant nuance and actually asked: “Is it more important to be right or to win money in your pool?” The difference is that if you only pick the favorites, your hit rate will be higher, but you won’t stand out from the other players in the pool. You have to bet on a few surprises to make money. I can put two brackets in my most important pool – the one where you get roundly teased if the bracket is bad – so I asked for one of each.
I followed ChatGPT’s recommendations almost completely. The exception was that I gave my three favorite teams a nod in the early rounds. Those teams are Santa Clara, Gonzaga and UCLA. I’d love to be wrong, but I don’t expect to see them later in the tournament.
My plan is to follow the success or failure of the ChatGPT predictions here as the tournament progresses. You’ll be the first to know if Chat GPT makes me look like an idiot. I keep my fingers crossed.
