People differ in their willingness to take risks. Recent work found that a dominant class of measures, revealed preference tasks (e.g., laboratory lotteries), appear not to tap into stable individual differences, whereas survey-based stated preferences are stable and predict real-world risk taking across different domains. How can stated preferences, often criticised as inconsequential ("cheap talk"), be more valid and predictive than controlled, incentivized lotteries? In our multi-method study, over 3,000 respondents from population samples answered a single widely used and predictive risk preference question. Respondents then explained the reasoning behind their answer. They tended to recount diagnostic behaviours and experiences, focusing on voluntary, consequential acts and experiences from which they seemed to infer their risk preference. We found that third-party readers of respondents' brief memories and explanations reached similar inferences about respondents' preferences, indicating the intersubjective validity of this information. Our results also shed light on the process of preference formation through experience over the lifespan. Finally, stated risk preferences may capture preferences revealed in behaviours in the wild better than the contrived behavioural tasks preferred in economics because they permit people to draw upon their own understanding of what constitutes diagnostic behaviours and experiences.