A vital application of game data mining is to predict player behaviour trends such as disengagement, purchase, etc.. Several works have been done by quantitative methods in the last decade. Generally, predicting player behaviour trends is a classification problem where class labels of instances are decided by predefined definitions. However, as the majority of current definitions distribute players into classes only by satisfying specific conditions, a highly biased class distribution may be led to if few (or most) players can satisfy these conditions. In this work, a new definition named trend over varying dates that can create balanced class distribution will be introduced and, as an example, disengagement prediction will be used to show how the definition works. Experiments on three commercial mobile games will show how this definition can be applied to games of various genres. Finally, the performance of this definition towards predicting disengagement will be compared with another disengagement concept called churn’. Both game-specific and event frequency based data representation (introduced in previous work) will be applied to represent the datasets for predictions. Results indicate that the definition of ‘trend over varying dates’ can improve the predictive performance by balancing the class distributions in most cases.