March madness prediction using machine learning techniques

Project Work presented as the partial requirement for obtaining a Master's degree in Information Management, specialization in Knowledge Management and Business IntelligenceMarch Madness describes the final tournament of the college basketball championship, considered by many as the biggest sporting event in the United States - moving every year tons of dollars in both bets and television. Besides that, there are 60 million Americans who fill out their tournament bracket every year, and anything is more likely than hit all 68 games. After collecting and transforming data from Sports-Reference.com, the experimental part consists of preprocess the data, evaluate the features to consider in the models and train the data. In this study, based on tournament data over the last 20 years, Machine Learning algorithms like Decision Trees Classifier, K-Nearest Neighbors Classifier, Stochastic Gradient Descent Classifier and others were applied to measure the accuracy of the predictions and to be compared with some benchmarks. Despite of the most important variables seemed to be those related to seeds, shooting and the number of participations in the tournament, it was not possible to define exactly which ones should be used in the modeling and all ended up being used. Regarding the results, when training the entire dataset, the accuracy ranges from 65 to 70%, where Support Vector Classification yields the best results. When compared with picking the highest seed, these results are slightly lower. On the other hand, when predicting the Tournament of 2017, the Support Vector Classification and the Multi-Layer Perceptron Classifier reach 85 and 79% of accuracy, respectively. In this sense, they surpass the previous benchmark and the most respected websites and statistics in the field. Given some existing constraints, it is quite possible that these results could be improved and deepened in other ways. Meanwhile, this project can be referenced and serve as a basis for the future work