This paper considers economic rationales for categorizing coarsely. Assuming that people make predictions about unobservable attributes in new situations they encounter on the basis of categories, we study the basic properties of alternative ways of categorizing in different environments. We show that coarse categorization may be optimal for making predictions in stochastic environments in which an individual has a limited number of past experiences. Furthermore, we find that the attempt to coordinate with others may be an additional rationale to categorize coarsely. In a complementary dynamic model we also analyse the adaptive behaviour of agents who have to learn to categorize in a dynamic setting.