Over the past two decades, respondents to the Shiller Investor Confidence Surveys have assessed the probability of a catastrophic stock market crash to be much higher than the historical frequency of such events. We decompose these crash probabilities into fundamental and subjective components and use a large language model to estimate the emotional content of respondent narratives. The subjective crash component is strongly associated with high negative affect. We use respondent location to test how news of unusual exogenous shocks affects crash belief formation. The results are consistent the risk-as-feelings hypothesis and suggest a path by which emotional response to news about salient events may play a role in the scale and variation in investor beliefs about rare disasters.