The Incredible Odds: Why AI can Never Pick the Perfect Bracket

Spoiler alert:  If you filled out a NCAA Basketball bracket it is busted.  So is mine. 

A few years ago, Google’s Alpha Go and its AI machine learning protégé Alpha Go Zero did something incredible.  They won a complicated game called Go against a world champion in a tournament.  It was another milestone for the promise of AI after previously crushing the human spirit in other games like Jeopardy and world champion Chess.  One thing though that AI will never do is pick a perfect NCAA basketball bracket.

There is a lot of data in basketball.  Sports in general have a lot of numbers thanks in part to new data collection methods and the growing field of Sports Analytics.  Basic basketball statistics are quickly being enhanced with passive player tracking data such as Second Spectrum , Shot Tracker, and Noahlytics.  All of these systems can capture 70+ data elements using visual machine learning and enhanced AI systems.  In short, the data is a whole lot more than free throw percentages and old-school shot charts.  But can more data translate in to better win predictions?

Against All Odds

Picking a winning team might seem easy.  Line up the stats – shooting percentages, offensive and defensive efficiency – and like magic (or a Logistic Regression Model with LASSO), it is simple to see Gonzaga had a 95% probability to win their first game.  The same approach suggested Kentucky was an 85% favorite over Saint Peters, but as fans witnessed the numbers (both the model and Kentucky’s scoring) fell short of that prediction.

Excluding the ‘early four’ games introduced in 2011, there are 64 teams to play 63 games in the tournament.  When you can easily see all the team match ups on one sheet of paper it seems deceivingly simple.  How hard could it be to pick the right combination of a complete bracket?  It turns out it is amazingly difficult.

Some fans claim they have a ‘method’, a way with the numbers, or perhaps secret strategy to pick the winners.  Best players, best record, primary colors and even toughest mascot have all been tried.  Completing a bracket using the coin-toss method results in 2^63 or 9.2 quintillion possibilities.  Yikes!  Experts might be able to narrow down the odds of picking a perfect bracket to 1 in 120 billion.  These odds are so outrageous that Warren Buffet once offered to pay $1 billon to anyone who predicted a perfect bracket. 

The staggering odds however are not what limits AI and machine learning.  It is not the machine it is the people.

The Hope and Dreams of AI 
There is a lot of promise for AI and what it can do and those lofty expectation just might be unrealistic.  AI has been great at solving some very specific computational but it is terrible at being human – or more poignantly anticipating humans’ actions.  This is why AI CAN predict the weather tomorrow (and reasonably well over the next five days) but CANNOT predict the stock market value one-day in advance. 

Humans, yes, us non-rational beings often mess up a perfect algorithm

Humans, yes, us non-rational beings often mess up a perfect algorithm by doing something that ‘just feels right’, is ‘on a whim’ as a result of a ‘gut feeling’ or living a ‘YOLO’ moment.  People often compare the human brain to some sort of super-computer with logical programs and paths for how the data is connected from inputs to outputs.  There is some truth to this and it might explain why customer marketing and recommendation engines like Amazon uses, are generally successful.  These work because predicting the behavior – the likes and dislikes – of a large group of people is possible.  Predicting one specific person’s actions in a particular moment is much more difficult.  (You don’t buy everything Amazon suggests, right? Right?) 

The point is, data used to forecast sports (or any other activity involving influential human behavior) can only go so far.  Because well, Humans mess things up.


What is the best AI can do?  Be more human, which is to say it needs to understand moments of irrational behavior and more importantly inspirational behavior.  When an athlete can believe more in themselves than the data does.  When they rise up after falling down to win.  When they make the impossible shot.  When they do something beyond the numbers, so extraordinary, even the algorithm pauses to take notice.  AI and Machine Learning are incredible, but so are humans in all of their infallibility and grace, and at times the wonder of their unpredictably.  It is what makes watching the tournament exciting and although I use data to fill out my bracket, I am rarely disappointed when someone does something unexpected and the underdog wins…

My opinions are clearly my own.

BONUS: When AI tries to be more human…

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s