The customer dropout is a phenomenon that often occurs in customers of sports services. This study intends to evaluate the performance of machine learning algorithms to predict dropout, using the available data of the customers and the history of the use of sports services to predict dropout. Several techniques were applied to perform the prediction to identify which was more accurate, as well the comparison of the accuracy of the prediction using a train / test and k-fold approach. The best performance was achieved with the algorithm Gradient Boosting Classifier in both approaches, although the runtime is high. The results show that no significative differences in the accuracy of the prediction and execution time of the techniques used in train / test and k-fold with p> 0.05 using Mann-Whitney.