Science of upsets: Prof has formula that doubles your shot

Message to hoops fans: This story could make you look brilliant.

A computer science professor at University of Illinois has created a formula that predicts NCAA tournament upsets at double the success rate of someone picking at random — including, but not limited to, those who throw darts at the bracket, or pick based on their favourite colour, the most ferocious mascot or the number of vowels in the coach’s last name.

This year’s upset picks both come out of the South region. They are No. 13 Buffalo over No. 4 Arizona, and No. 14 Wright State over No. 3 Tennessee.

But before betting the mortgage, read on:

The computer scientist who spearheads this project, Sheldon Jacobson, says the computer models only analyze potential upsets by 13, 14 and 15 seeds.

Jacobson and fellow scientists pared down 115 publicly available metrics for every team in college basketball to 15 that have served as the best predictors of upsets in years past.

Some examples include effective possession ratio — essentially the number of points a team scores per possession — along with average scoring margin and opponent’s 3-point shooting percentage.

Now for the science: The framework of these formulas is called “balance optimization subset selection” (BOSS), which is an artificial-intelligence algorithm (Google that if so inclined).

March Madness generates more than $10 billion a year in wagering, much of which comes when players chip in $10 or $20 and fill out brackets for their office pools and collect points based on the number of correct picks.

Picking the eventual champion — No. 1 seeds Villanova and Virginia started at 5-1 odds to win it all, with No. 2 Duke at 6-1 — always helps. But sometimes the real difference makers are the correct upset picks in the early rounds. That’s when the Buffalos and Wright States of the world beat Goliath and briefly restore faith in the gumdrops-and-lollipops notion that anything really is possible.

It’s not, Jacobson assures us. Still, his website, bracketodds.cs.illinois.edu, gets thousands of hits a day this time of year.

Once the brackets were revealed, Jacobson set the computer’s focus toward picking these upsets. Its track record since 2003 is hardly perfect, but still probably better than yours.

Using BOSS, the computer picks the two most likely upsets each year. Last season, not a single 13, 14 or 15 advanced, so it got 0 per cent. Two seasons ago, there were three such upsets — Iona and Buffalo — but the computer didn’t pick either of them.

But in 2015, BOSS picked Georgia State and UAB and went 2 for 2.

And since 2003, 10 of its 26 selected games have resulted in upsets. That’s 38.4 per cent, or double the expected number of correct selections a person would get by using a “weighted random selection method.” In other words, double what you’d get by choosing a team because you like the fight song.

For those placing faith in his science, Jacobson warns of the large gulf between predicting the future and forecasting what could happen.“Nobody predicts the weather. They forecast it using chances and odds.”

Similarly, he says, “artificial intelligence looks at some outcomes that the human eye can’t catch. The models we use give some indication of what the future may look like.”