Sales prediction is a complex task because of a large number of factors affecting the demand. We present a context aware sales prediction approach, which selects the base predictor depending on the structural properties of the historical sales. In the experimental part we show that there exist product subsets on which, using this strategy, it is possible to outperform naive methods. We also show the dependencies between product categorization accuracies and sales prediction accuracies. A case study of a food wholesaler indicates that moving average prediction can be outperformed by intelligent methods, if proper categorization is in place, which appears to be a difficult task.