This talk is concerned with two settings for relational learning and their connection to the 'reference class problem'. In the first setting one aims to learn first-order generalisations of relational data by abstracting over individuals. In the second, one explicitly represents and approximates hidden properties of individuals. Under explicit assumptions about the underlying generative process, we show why models which abstract over individuals cannot achieve optimal prediction. We provide empirical evidence demonstrating that models which capture hidden properties of individuals achieve greater accuracy, and discuss the potential of such models for achieving optimal prediction.