University Of Wisconsin Madison Prediction of NBA games

University Of Wisconsin Madison Prediction of NBA games based on Machine Learning Methods ECE/CS/ME 539 Introduction to Artificial Neural Networks and Fuzzy System Renato Amorim Torres December, 2013

Proposal The objectives of the project are: Predict the winning team of a NBA game The goal is to have the Prediction Rate higher than the rate of the Very Naive Majority Vote Classifier. This method looks at all previous games (in the season) of the two teams and picks the team with the fewest losses as the winner.

Preparation of the Data was obtained on the website: www. basketball-reference. com

Preparation of the Data Box scores were copied into a spreadsheet Using macro all the team names were replaced by numbers, the unnecessary columns were deleted, and. txt files were generated with the data. The. txt files were loaded in Mat. Lab and used to generate the feature vectors and implement the methods.

Data Analysis Blue Line >> 2012 to 2006 Regular Seasons Red Line >> Mean of

Feature Vectors Win-Loss percentage of both teams Point differential per game of both teams Win-Loss percentage in the previous 8 games for both teams Visitor Team win-Loss percentage as visitor Home Team win-Loss percentage at home Total of eight features

Feature Vectors Win-Loss percentage in the previous 8 games of both teams. Why eight games? Blue Line >> 2012 to 2006 Regular Seasons Red Line >> Mean of

Prediction Based on the graph shown before, it was found that each team should have played at least eight games prior to predicting games. Considering this starting point, the Naive Majority Vote Rate was calculated in order to define the goal: Naive Majority Vote Rate from 9 th game Data Rate(%) 2012 61. 83 2011 2010 2009 2008 2007 2006 Mean 65. 94 64. 75 66. 30 63. 86 65. 81

Prediction Maximum Likelihood Classifier - Not all the feature vectors were used. In order to establish the best features, a code was implemented to verify the prediction rate for all combinations of features. Features used: Feature #2: Home Team Win-Loss percentage at home Feature #4: Home Team total Win-Loss percentage Feature #5: Visitor Team point differential per game in the actual season Feature #7/8: Win-Loss percentage in the last 8 games for both teams

Prediction Maximum Likelihood Classifier -The training data was a combination of the previous season's games and the testing data was random games of the “current” season. The likelihood classifier achieved the following results: Likelihood Classifier Testing Data 2012 2011 2010 2009 2008 2007 2006 Mean Rate(%) 67. 31 64. 37 67. 88 67. 43 69. 43 65. 05 67. 42

Prediction Maximum Likelihood Classifier Results Likelihood Classifier Testing Data 2012 2011 2010 2009 2008 2007 2006 Mean Rate(%) 67. 31 64. 37 67. 88 67. 43 69. 43 65. 05 67. 42 Naive Majority Vote Rate from 8 th game Data 2012 Rate(%) 61. 83 2011 2010 2009 2008 2007 2006 Mean 65. 94 64. 75 66. 30 63. 86 65. 81 For comparison. . . Table extracted from: Matthew Beckler, Hongfei Wang. and Michael Papamichael NBA Oracle

Prediction Linear Regression / LMS Algorithm - In order to find the weights, the LMS Algorithm was implemented: