Predicting NHL Match Outcomes with ML Models

Abstract

Finding ways of predicting the outcome of sports games from performance data has always been an attractive proposition for many statisticians and, lately, of data miners using machine learning (ML) techniques. A research paper [1] on ice hockey (NHL) by a University of Ottawa team (from now on referred simply as Ottawa), and their generous sharing of the data used for their research provided the main drive for this paper. In this research, the Ottawa data is used for a number of purpose, all involving the use of ML techniques to predict the outcome of NHL games. First, we repeat Ottawa's experiment, which looked " ...at how effective traditional, advanced and mixed (both combined) statistics were for predicting success in the NHL " . Then we split all the given attributes in the data into Categorical and Continuous, and build ML separate models, whose result we compare with those of the original model. The original data is also parsed to create a new dataset, and models built to compare with the results of the original one. Lastly, a framework for making use of this data in a practical application (betting) is proposed, and the accuracy of models built is evaluate and compared. Three ML techniques: Decision Trees (DT), Artificial Neural Networks (ANN), and ClusteR, a software developed by a betting company, were used for these experiments.