Illustration for article titled How to Use AI to Win March Madness (maybe)Photo: Justin Casterline (Getty Images)

For most people, predicting the results of the NCAA March Madness tournament with some degree of accuracy is a breeze. The singles elimination tournament, made up of 64 teams, one of which must win six games in a row to be crowned champion, is notoriously unpredictable. Unknown Davids disturb the ancient Goliaths almost every year.

Enter the machines. Given the popularity of prediction algorithms and data models influencing elections over the past few years, it’s no surprise that March Madness has gone mad with a barrage of AIs trying to predict the big dance. Knowing this, perhaps while staring at a daunting bracket, the question arises: should you consult an AI before filling out your bracket?

The answer, which depends on who you are and how you understand the intricacies of college basketball, may be clear.

Illustration for article titled How to Use AI to Win March Madness (maybe)

Can AIs predict the tournament better?

According to Matt Osborne, mathematician and postdoctoral fellow at Ohio State University, an AI-generated model can help better inform your parenthesis in three ways.

G / O Media can receive a commission

As he explains to Lifehacker, a predictive model can give you a good idea of ​​where things are likely to go in the tournament, but it certainly can’t predict the tournament directly. He explains that there are certain criteria that can increase your chances of accuracy when using an AI: overcoming your personal biases (which are widespread in sports), measuring historical results, and making a general estimate of likelihood (which he did with the traditional Las Vegas compares) betting odds).

In an email, he elaborates on the following:

You may have personal biases that are not present in the data. For example, you think Team A is good because you’ve heard of them before, but it turns out they aren’t that great this year.

Using historical results can help you guess better than random. Choosing teams at random is basically like flipping a quarter, but the chances of Team A beating Team B are usually not 50-50. For example, a 16-seed only upset a 1-seed once in tournament history (UMBC v Virginia in 2018) so you’re pretty much sure to tag the top dogs in each region.

[Algorithms] can give you an estimate of the likelihood of something going to happen (similar to Vegas odds) which can help you determine for yourself how likely it is that you are contemplating an angry choice.

According to Osborne, AI tools are generally useful for amateur fans and well-trained experts alike. However, he adds that die-hard fans “are likely to be better equipped (based on their existing knowledge) to understand the input and output of the tool and see if what the machine is suggesting actually sounds feasible.”

Illustration for article titled How to Use AI to Win March Madness (maybe)

How accurate are the AI ​​brackets usually?

Asking an AI to predict a college basketball tournament is a little thornier than asking one to predict a presidential election. Although the latter – as in 2016 – may be subject to irregularities, the models are fed a steady stream of surveys based on the more reliable act of voting. Basketball is more volatile, especially in an environment like March Madness.

In Osborne’s experience, “Only the best brackets are ever covered. Since the tournament is pretty random, the absolute best brackets are usually not created with a data science tool.” As an example, he refers to the case of 12-year-old Sam Holtz, who filled in a perfect bracket at the ESPN March Madness Challenge 2015 and defeated over 11 million other participants. Crazy enough, Holtz didn’t even watch basketball regularly and operated on simple premonitions when choosing his winners. And then he kind of made history, filling in a perfect bracket that defied the gigantic odds that an economics professor calculated, “somewhere in one of a few trillion.”

No predictive model has ever done this, and it is unlikely that it will. But that doesn’t mean that an AI-generated bracket isn’t useful if you decide to enter a competition. Osborne recommends for signposts The model from Five Thirty Eight (which is free) and The model from Sportsline (which comes with a fee).

Even with a model available, it should also make you feel instinctive and biased. After all, we’re talking about sports.