Basketball fans have one more day to fill out their March Madness brackets. They’ll need to pick not just the champions and their route to victory, but predict the paths of all the losers as well. It’s not easy. In fact, no person or computer has yet been able to do it.
But that’s not for lack of trying. Andy Dieckhoff has been working on his NCAA basketball ranking system for more than a decade. He’s a senior at Portland State University studying applied linguistics. But in his spare time he studies basketball.
“I get really excited,” he says. “I tend to think of Bracket Day as more enjoyable than my birthday most years.”
Dieckhoff started watching games with his dad when he was 5 or 6, and he ran his first bracket pool in middle school.
“I wanted to give myself kind of an advantage,” he says.
So he poured a bunch of win-loss stats into a spreadsheet and used some basic equations to try and predict the winners.
“It was nothing sophisticated or anything like that when it started,” Dieckhoff says.
But since that first year, when Dieckhoff beat all the 11-year-olds in his pool, his system has become a lot more complex. The spreadsheet has grown to include not just scores, but rebounds, assists, three-point shooting percentages and numerical values for some qualities that are harder to define, like hustle, discipline and even luck. The system is all about using intuition to interpret the stats.
“The weighting of everything is done on completely gut feeling,” Deickhoff says. “I’ve picked up things I feel are important in the game.”
Dieckhoff’s system correctly predicted 66 of the 68 teams that made it into this year’s tournament. But that doesn’t necessarily mean it will help him predict winners in the tournament.
“There’s no making sense of it,” Dieckhoff says. “I’m trying pretty hard to make sense of it. But at the end of the day, any team can win any game.”
Dieckhoff isn’t alone. More than 3 million people filled out brackets on ESPN.com last year.
Keith Lipscomb, a senior editor at ESPN Fantasy Games says no one got a perfect bracket; the best anyone did was guessing 52 out of the 64 games. Raw knowledge of basketball doesn’t count for that much when it comes to filling out brackets.
“The NCAA tournament happens to be my favorite thing in the world,” Lipscomb says. “But that doesn’t mean I have any better grip on it then anybody else just because I watch a lot of games. I feel like you’re sometimes better off not knowing too much, because that way you don’t overthink it.”
So what if you know nothing? What are the odds of randomly predicting the outcome of every tournament game?
According to Mike Weimerskirch, a math professor and sports fan at the University of Minnesota, those odds aren’t good: about 147 quintillion to one (that’s 147,000,000,000,000,000:1). The odds get slightly better (about 9 trillion to one) if you ignore the play-in games and just look at the field of 64.
“It’s going to be more likely for Phil Mickelson to get a hole in one on all four of the par threes in the opening round of the Masters than it is to fill out a perfect bracket,” Weimerskirch says.
But what if you take a slightly smarter approach to filling out your bracket and always pick the teams that are seeded higher?
“You bring it down to about 150 billion to one,” Weimerskirch says.
But in the age of Google, shouldn’t we be able to get closer than that? Shouldn’t we be able to use all our computational power to correctly predict all the winners and create the perfect bracket?
That’s the question behind a very special March Madness pool run by Danny Tarlow, a postdoctoral student at Microsoft Research Cambridge. His pool has just one serious rule: no humans.
Computer programs fill out the brackets.
It started a few years ago when Tarlow entered a March Madness bracket and wanted to give himself a bit of an edge. He hijacked the basic structure of a program he was working on (it was designed to anticipate a reader’s book selection) and used it to predict winners.
“It worked surprisingly well,” Tarlow says. “I won the pool, so it wasn’t behaving completely crazily.”
The next year he invited other programmers to enter their own bracket-building algorithms.
Each competitor enters a program that must learn about basketball by chewing through the stats from past games. It’s similar to voice recognition software learning speech. The programs guess the outcome of a game, then refine their algorithms based on the actual results. Then they repeat the process, guessing and checking, with all the games on record.
“It can take quite a long time to do this initial learning phase — several hours to even a day,” Tarlow says.
So are these programs any good at filling out brackets?
“Looking at the group of algorithms, it’s probably not that much different than you would expect to see out of your group of friends,” Tarlow says.
It turns out that the cold statistical precision of a computer program is just as unsuccessful as human intuition. Weimerskirch says there just isn’t enough data to overcome the randomness of college basketball.
“No matter how much computer analysis you do, you’re still stuck with the way the ball bounces,” he says.
“That’s the beauty of college basketball,” Dieckhoff says. “There are upsets all the time. Maybe sometimes it’s better to just put on a blindfold and pick the teams.”
Or you could try a more creative strategy.
“Somebody apparently won their office pool basing their picks on who would win the game if the mascots fought,” Weimerskirch says.
In some years that could bring up some tricky questions.
That kind of unpredictable match-up is what makes March Madness so … maddening.