Predictions of european basketball match results with machine learning algorithms

The goal of this paper is to build and compare methods for the prediction of the final outcomes of basketball games. In this study, we analyzed data from four different European tournaments: Euroleague, Eurocup, Greek Basket League and Spanish Liga ACB. The data-set consists of information collected from box scores of 5214 games for the period of 2013-2018. The predictions obtained by our implemented methods and models were compared with a vanilla model using only the team-name information of each game. In our analysis, we have included new performance indicators constructed by using historical statistics, key performance indicators and measurements from three rating systems (Elo, PageRank, pirating). For these three rating systems and every tournament under consideration, we tune the rating system parameters using specific training data-sets. These new game features are improving our predictions efficiently and can be easily obtained in any basketball league. Our predictions were obtained by implementing three different statistics and machine learning algorithms: logistic regression, random forest, and extreme gradient boosting trees. Moreover, we report predictions based on the combination of these algorithms (ensemble learning). We evaluate our predictions using three predictive measures: Brier Score, accuracy and F-1-score. In addition, we evaluate the performance of our algorithms with three different prediction scenarios (full-season, mid-season, and play-offs predictive evaluation). For the mid-season and the play-offs scenarios, we further explore whether incorporating additional results from previous seasons in the learning data-set enhances the predictive performance of the implemented models and algorithms. Concerning the results, there is no clear winner between the machine learning algorithms since they provide identical predictions with small differences. However, models with predictors suggested in this paper out-perform the vanilla model by 3-5% in terms of accuracy. Another conclusion from our results for the play-offs scenarios is that it is not necessary to embed outcomes from previous seasons in our training data-set. Using data from the current season, most of the time, leads to efficient, accurate parameter learning and well-behaved prediction models. Moreover, the Greek league is the least balanced tournament in terms of competitiveness since all our models achieve high predictive accuracy (78%, on the best-performing model). The second less balanced league is the Spanish one with accuracy reaching 72% while for the two European tournaments the prediction accuracy is considerably lower (about 69%). Finally, we present the most important features by counting the percentage of appearance in every machine learning algorithm for every one of the three analyses. From this analysis, we may conclude that the best predictors are the rating systems (pi-rating, PageRank, and ELO) and the current form performance indicators (e.g., the two most frequent ones are the game score of Hollinger and the floor impact counter).

Automated summary

The goal of this study is to develop and compare methods for predicting the final outcomes of basketball games. By analyzing data from four European tournaments, the researchers found that new performance indicators and historical statistics can improve predictions significantly, and these indicators can be easily obtained in any basketball league. By using various statistical and machine learning algorithms, such as logistic regression, random forest, and extreme gradient boosting trees, the researchers achieved higher accuracy compared to the traditional model, with a 3-5% improvement.