Predicting Outcomes of NBA Basketball Games

Abstract

A stratified random sample of 144 NBA basketball games was taken over a three-year period, between 2008 and 2011. Models were developed to predict point spread and to estimate the probability of a specific team winning based on various in-game statistics. Statistics significant in the model were field-goal shooting percentage, three-point shooting percentage, free-throw shooting percentage, offensive rebounds, assists, turnovers, and free-throws attempted. Models were verified using exact in-game statistics for a random sample of 50 NBA games taken during the 2011-2012 season with 88-94% accuracy. Three methods were used to estimate in-game statistics of future games so that the models could be used to predict a winner in games played by Team A and Team B. Models using these methods had accuracies of approximately 62%. Seasonal averages for these in-game statistics were used in the model developed to predict the winner of each game for the 2013-2016 NBA Championships.