Neural Networks may become next best way to predict NFL Games
Neural Networks may become next best way to predict NFL Games
Student researchers at The College of Wooster developing ways to outsmart prognosticators
Contact
John Finn
330-263-2145
Email
WOOSTER, Ohio - By the time the Browns take on the Steelers in week six of the upcoming NFL season, Saif Ahmad and Michael Janning believe they'll have a pretty good idea who will win. It's not that they are expert football prognosticators - in fact, Ahmad is from Jamaica and has never seen an NFL game in person - but as students in The College of Wooster's Applied Mathematics and Research Experience (AMRE), they are finding ways to build mathematical models and identify critical statistics that provide the most reliability in predicting outcomes.
By using artificial neural networks (algorithms that learn patterns from data, similar to the way your brain works), the students are looking at the most dependable indicators of outcome. "What you have is a bunch of neurons that interact with each other," said John David, visiting assistant
professor of mathematics and computer science at Wooster and one of the advisers for the project. "The more you do something, the more certain neural pathways get built up, enabling you to better predict similar future situations. For example, because offensive linemen typically weigh more than
wide receivers, an algorithm would process the variable (weight) and predict the position of the player based solely on the input. Our system takes the emotion out of the prediction. This system is based solely on numbers, and arguably the most objective way to predict the outcome of NFL games."
As most football fans might surmise, the combination of one team's passing offense compared to the other team's passing defense is a key factor in the outcome of a game. Likewise, rushing offense versus rushing defense is also a critical determinant. In addition, turnovers have been identified as significant variables, with interceptions appearing to matter more than fumbles. Obviously, point differential and team records are also noteworthy considerations. Overall David and his partner on the project, Drew Pasteur, assistant professor of mathematics and computer science at Wooster, have settled on 11 statistical comparisons per team, plus home-field advantage for a total of 23 variables. "What really matters is how the two teams match up," said David. "Say one team has a great passing offense and its opponent has a weak secondary, while the second team has a good offensive line, and the first team has a week defensive line. What happens? How will that affect the outcome? Our model is designed to learn to predict that."
While these variables may seem obvious to the average fan, they are not the focus of the study. "What we are really trying to understand, and what inspired us to pursue this, are the underlying networks and how they function," said David, who is advising the project with Pasteur."We are learning as much about the process as we are about predicting outcomes of the games. The same methodology was also used for a similar project we did for Goodyear to predict tire performance based on the rubber used to manufacture them."
David compared the neural networks to the models that Netflix and Amazon.com use to predict consumer interest and potential purchases. "More and more companies are using this approach to predict future outcomes," he said. "Any time a website makes a recommendation based on previous behavior, it is employing the same principles we used for this project."
Accessing the statistics is not difficult - the information is readily available online at such sites as ESPN.com - but deciding which ones to use is critical. "Introducing more variables does not necessarily guarantee a more accurate prediction," cautions Pasteur, who also runs a highly successful website (Fantastic50.net) that predicts the outcome of high school football games and
which teams are most likely to make the playoffs. "In fact, it could become more complicated because even a slight change in input could result in a dramatic change in output."
Ahmad and Janning selected variables from the 2007 and 2008 seasons, and were able to make fairly accurate predictions about the 2009 season. "We were right with the experts," said Ahmad. "Our predictions have been correct around two thirds of the time in terms of accuracy, which is comparable to the experts at ESPN.com, but we hope to build an even more accurate model by determining the most reliable variables."
As the season progresses, the predictions should become even more accurate because the neural network will become more refined. As for the Browns and Steelers on Oct. 17? Well, based on last year's numbers when the two teams split the regular-season series, Ahmad and Janning say the Steelers would be favored by 9.66 points. Of course things may change dramatically when the two teams meet again on Jan. 2.
And the Super Bowl? Based on simulation of the actual 2010 season and playoffs (using 2009 statistics) the Jets will beat the Eagles and claim their first title since 1969.