A Machine Learning Approach to March Madness∗

The aim of this experiment was to learn which learning model, feature selection technique, ordering process, and distance criteria provided the best classification accuracy in predicting the outcome of any NCAA Men’s Basketball Tournament match. Ninety-four features were selected from ESPN.com for each team accepted into the tournament for the last four years. Ordering processes were tested against the baseline and the random ordering increased classification accuracy from 0.61 to 0.63. Random ordering was used to test a variety of feature reduction techniques. Random forest feature reduction performed best and increased accuracy from 0.63 to 0.7322 when used in conjunction with kNN and limiting the number of features to five. Using Manahattan distances as opposed to Euclidean distance further increased accuracy from 0.7322 to 0.7362.