Churn describes customer defection from a service provider. This can be observed in online freemium games, where users can leave without further notice. Game companies are looking for methods to detect and predict churn to enable management reaction. The recorded data of games can be analyzed for this purpose. In this article, we conducted a case study based on data from the freemium game The Settlers Online. Churn detection was achieved by application of four different labeling approaches, based on common churn and disengagement definitions within the game analytics literature. In order to model predictive classifiers, features were computed from the raw game data. Eight different machine learning algorithms returning binary classifications were applied. The results were compared for all algorithms regarding all labeling approaches. Random forests with sliding windows were the best solution in our case, returning area under curve values higher than 0.99, thereby enabling prediction accuracies of 97% in our dataset. The results were confirmed by tests on an independent dataset and in our discussion, we offer guidance on the interplay of feature engineering, labeling approaches-in particular, disengagement-and machine learning algorithms for churn prediction. Our recommendations are valuable for game companies and academics, who pursue similar studies.