The most common approach to assessing natural hazard risk is investigating the willingness to pay in the presence or absence of such risk. In this work, we propose a new, machine-learning-based, indirect approach to the problem, i.e. through residential-choice modelling. Especially in urban environments, exposure and vulnerability are highly dynamic risk components, both being shaped by a complex and continuous reorganization and redistribution of assets within the urban space, including the (re-)location of urban dwellers. By modelling residential-choice behaviour in the city of Leipzig, Germany, we seek to examine how exposure and vulnerabilities are shaped by the residential-location-choice process. The proposed approach reveals hot spots and cold spots of residential choice for distinct socioeconomic groups exhibiting heterogeneous preferences. We discuss the relationship between observed patterns and disaster risk through the lens of exposure and vulnerability, as well as links to urban planning, and explore how the proposed methodology may contribute to predicting future trends in exposure, vulnerability, and risk through this analytical focus. Avenues for future research include the operational strengthening of these linkages for more effective disaster risk management.