Churn describes customer defection from a service provider.This can be observed in online freemium games, where userscan leave without further notice. Game companies are lookingfor methods to detect and predict churn to enable managementreaction. The recorded data of games can be analyzed forthis purpose. We conducted a case study based on data fromthe freemium game The Settlers Online. Churn detection wasachieved by application of four different labeling approaches,based on common churn and disengagement definitions within thegame analytics literature. In order to model predictive classifiers,features were computed from the raw game data. Eight differentmachine learning algorithms returning binary classifications wereapplied. The results were compared for all algorithms regardingall labeling approaches. Random forests with sliding windowswere the best solution in our case, returning AUC valueshigher than 0.99, thereby enabling prediction accuracies of 97%in our data set. The results were confirmed by tests on anindependent data set and in our discussion, we offer guidanceon the interplay of feature engineering, labeling approaches-inparticular disengagement-and machine learning algorithms forchurn prediction. Our recommendations are valuable for gamecompanies and academics, who pursue similar studies.