Predict To Succeed : Optimal Sequential Fantasy Football Squad Formation Using Machine Learning Tools

Fantasy football is a popular game in which participants assemble a roster of real-world athletes and gain points for their successful performance in football matches. Fantasy sports appeared as a unique phenomenon in the USA in 1980s [2]. The rapid evolution of the Internet turned fantasy sports into an easily accessible entertainment for millions of people worldwide. Fantasy games are usually timed to coincide with major championships and events, either single (i.e. the FIFA World Cup) or season-long (e.g. the English Premier League). The most popular fantasy sport subject in Europe is association football. Fantasy tournaments are often run either by media (e.g. www.sports.ru/fantasy) or by leagues’ organizers (the Fantasy Premier League [1], the Official Bundesliga Fantasy). In this paper, a model of optimal sequential decision making in the Fantasy Premier League (FPL) is presented. There were more than 4.5 million participants in the 2016/2017 FPL season. This fact makes the FPL the most competitive fantasy league in the world. A high level of competitive intensity requires making the best decisions over time in order to succeed in the FPL. This task is particularly demanding as there is a spanless set of possible actions, moreover, the actions should be made under uncertainty. Thus, there is a need for a model that will be powerful enough to perform well under these circumstances. The general idea is to maximize the number of points scored in a given round by solving an integer programming problem such that an objective function sums number of points predicted by a machine learning algorithm and constraints describe corresponding game rules. To predict expected points, we apply XGBoost, one of the most effective modern data science methods. We then build an automated manager that utilizes this approach and achieves rather decent results in team formation.