Predicting Upsets

In one of the first upsets of 2010's NCAA Men's Basketball Tournament, Old Dominion (an eleven seed) defeated Notre Dame (a six seed) by a single point, 51-50. Fans everywhere stared dismally at the brackets they had painstakingly filled out for their betting pools. How could they have foreseen such an unexpected outcome? Turns out, a college basketball ranking system based on a combined logistic regression/Markov chain (LRMC) model was able to, and so precisely that it predicted the one-point difference in the score.¹

LRMC was first developed by Joel Sokol and Paul Kvam. LRMC uses mathematical analysis to determine how far a team would advance in the tournament based on its performance and its strength of schedule throughout the preceding season. The model is similar to other formulas widely used to rank the 65 teams in the tourney. LRMC uses basic input data, such as the location, winner, and score of each season game, to make its predictions. However, this model analyzes how much playing at home actually helps a team win rather than assuming that every home game is worth the same to every team. In addition, it devalues close games. Based on statistical research, the developers of LRMC observed that close games went to the better ranked team a little more than half the time. So this model's ranking methodology bases its calculations on the principle that winning a close game does not elevate a team's standing as much as blowing out another team would.²

Over the past seven years, LRMC has been shown to be statistically more accurate at predicting tournament outcomes than other sources fans usually rely on. These include polls conducted by ESPN and Associated Press sportswriters, and the Ratings Percentage Index (RPI), a formula that also ranks the March Madness teams based on their wins, losses, and strength of playing schedule during the regular season.²

So next time you're trying to decide which double-digit seed will make it to the Final Four, look into using LRMC as a predictor. Trusting the math seems safer (for your pride and your wallet) than a decision based on another tried-and-true method, choosing which school mascot reigns supreme (in which case I'm still rooting for Old Dominion's Monarchs over Notre Dame's Fighting Irish)!

Image Credit: http://wikimedia.org

References:

1. O'Hara, C. "NCAA Bracketology: The Science Behind March Madness." PBS Newshour. March 18, 2010.

2.  Kvam, P. & Sokol, J. S.  A logistic regression/Markov chain model for NCAA basketball. Naval Research Logistics 53, 788–803 (2006).