Various methods based on classical classification methods such as linear discriminant analysis (LDA) have been developed for working on data streams in situations with concept drift. Nevertheless, the updated classifiers of such methods may result in a bad prediction error rate in case the underlying distribution incrementally changes further on. Therefore, we invented a rather general extension to such methods to improve the forecasting quality. Under some assumptions we estimate a model for the time-dependent concept drift that is used to predict the forthcoming distributions of the features. These predictions of distributions are finally used in the LDA to build the classification rule and hence for predicting new observations. In a simulation study we consider different kinds of concept drift and compare the new extended methods with the methods these are based on.