Canonical work on human decision-making demonstrates consistent, stereotyped and sub-optimal risk-reward trade-offs. However, recent work on experienced-based-, perceptual- and sensori-motor choice appear to severely restrict the scope of the canonical work, demonstrating trade-offs that are either differently sub-optimal, or much closer to optimal. Such dissociations may reflect highly specific mechanisms, or other confounds between domains; impossible to tease apart with current observational methods. We develop a method for strong causal inference based on experimentally manipulating risk preferences. Using a double-blind randomized control design, we trained people in a single domain, evaluate the extent to which training generalizes beyond the training-set, and more importantly the extent to which training transfers to risk domains in which participants were not trained. Participants were trained to maximize their expected earnings (i.e., to be risk-neutral). In total, we tested for transfer to four different risk domains, all of which are known to dissociate. We find that training is necessary and sufficient to cause reliable changes in decision-making. Importantly, training transfers, with participants becoming more risk-neutral also in domains for which they had had no training, showing that risk-reward trade-offs in different domains and tasks share common substrate. Shared mechanisms open up the opportunity to for practical general-purpose training programmes to improve human decision-making.