Georgia Tech Computer System Predicts NCAA Basketball Tournament Picks
Posted March 14, 2011 | Atlanta, GA
You don’t need a crystal ball to win your NCAA basketball March Madness bracket pool this year.
Just check out the predictions by LRMC (Logistic Regression Markov Chain), the computer ranking system designed by three professors at the Georgia Institute of Technology.
It predicts this year’s NCAA Final Four matchups will most likely be Ohio State vs. Duke and Kansas vs. Brigham Young University (BYU), with Ohio State beating Kansas for the championship.
The southeast region has the biggest likelihood of first-round upsets by double-digit seeds, with Michigan State, Gonzaga, Utah State and Belmont looking like all good candidates.
“Since one of BYU’s best players left the team, the southeast region is really wide open,” said Joel Sokol, operations research professor in the Stewart School of Industrial and Systems Engineering at Georgia Tech.
Since the 2000 season, LRMC has correctly predicted the outcomes of more NCAA tournament games than competing ranking systems and major polls.
Last year, LRMC correctly predicted the winners of 51 out of 64 NCAA games – beating out more than 50 of the top ranking sites. In 2008, the system predicted not only the Final Four, final two and the eventual victor, but also several upsets in earlier rounds.
Developed by Georgia Tech Professors Joel Sokol, Paul Kvam and George Nemhauser, with assistance from Mark Brown, math professor at City College of New York, LRMC utilizes data such as home court advantage, scores, teams competing and margin of victory in past performances to calculate likely victors.
But it also uses statistical methods to determine potential underdogs, which Kvam calls the “six degrees of Kevin Bacon approach.”
“The system bounces from team to team looking at their results, trying to hone in on who is really No. 1,” Sokol said. “The team that it keeps coming back to most often is our No. 1, and so on.”
With 68 teams vying for college basketball's biggest prize, more teams than ever before, the tournament can always be affected by upset, injuries or last-second, buzzer-beating baskets. That’s the human factor where LRMC predictions can falter.
But the system has proven more reliable with its predictions than the NCAA’s own Ratings Percentage Index (RPI). Historically, the upgraded LRMC method has picked the winner of more than 74 percent of tournament games correctly, while the RPI has been right less than 70 percent of the time.
Sokol recommends starting with LRMC predictions and making tweaks based on your own personal knowledge or preference.
“It’s like Watson,” Sokol said of the IBM Supercomputer that appeared on “Jeopardy!” last month. “Overall LRMC is likely to be better than others, but every once in a while it says something that you look at and say, ‘How could that be?’”