Predicting NFL Game Outcomes: Back-Propagating MLP By Paul McBride
Project Goal To predict the outcome of NFL games. Remove human bias Create a completely objective and statistics based prediction method
Why a back-propagating MLP? Since there are many ways a team can win, no linear mapping exists to conclude the outcome of the game This is a pattern classification problem
Data Collection I collected my data from NFL.com I chose to look at the entire 2012 season Since the NFL is an incredibly offense dominated league, I decided to compare offenses
Statistics 15 stats: Homefield, Firstdowns Totals, Totals yards, PassYards, etc. Extracted a feature vector for each game played by taking the differential statistics of offensive performance.
Example Feature Vector Team 1 vs. Team2: Each feature = Team 1 stat – team 2 stat Outcome of 1 = Team 1 won. Outcom of -1 = Team 2 won.
Support Vector Machine 4 – Way Cross validation. Linear kernel function with C = 1 proved to be a good result Confusion Matrix Classification rate of 0.887795276 225 29 28 226
MLP Preprocessed the data with SVD 4-Way cross validation to decide best classification rate 3 layers hidden layer neurons = 5 mu = .2, alpha = .007 - Classification rate = 88.1234%
Predicted Week 15, 2013 SVM: MLP: Chargers Broncos Redskins Falcons Buccs Seahawks Giants Eagles Vikings Pats Dolphins Bills Jags Texans Colts Bears Browns Chiefs Raiders Jets Panthers Packers Cowboys Cards Titans Saints Rams Bengals Steelers Ravens Lions Chargers Broncos Redskins Falcons 49ers Buccs Seahawks Giants Eagles Vikings Pats Dolphins Bills Jags Texans Colts Bears Browns Chiefs Raiders Jets Panthers Packers Cowboys Cards Titans Saints Rams Bengals Steelers Ravens Lions
Future I would like to trim down some of the less performance indicative stats I would like to add defense