In-Play Outcome Prediction of an Ongoing Basketball Game Using Machine Learning

Abstract

Basketball is among the most popular sports in the world today. Leagues and tournaments are played in all regions of the world, for example, the NBA, EuroLeague etc. Because of huge interest in basketball, there is a sizable pool of followers who are interested in betting. We study the related problem of predicting the winner of an ongoing basketball game which revises as the match progresses (also referred to as in-play winner prediction by researchers). We propose a supervised machine learning framework for the in-play winner prediction problem in basketball which seems to be new in the literature. We also quantify the importance of various features in arriving at the predictions using explainable machine learning techniques. Specifically, we use the Shapley Additive Explanations (SHAP) framework. We evaluate our approach on a dataset of regular season NBA games from 8 seasons (2015–16 to 2022–23 season). Our framework achieves the best overall prediction accuracy of about 74% over all states of the game, and the best in-play prediction accuracy ranging from about 63% to about 90% depending on the state of the game. SHAP scores reveal that different classification algorithms learn to predict differently. That is, feature importance (i) depends on different classification algorithms, and (ii) varies with time.