The financial press is a conduit for popular narratives that reflect collective memory about historical events. Some collective memories relate to major stock market crashes, and investors may rely on associated narratives, or crash narratives, to inform their current beliefs and choices. Using recent advances in computational linguistics, we develop a higher-order measure of narrativity based on newspaper articles that appear following major crashes. We provide evidence that crash narratives propagate broadly once they appear in news articles and significantly explain predictive variation in market volatility. We exploit investor heterogeneity using survey data to distinguish the effects of narrativity and fundamental conditions and find consistent evidence. Finally, we develop a measure of pure narrativity to examine when the financial press is more likely to employ narratives.

We document that variation in the question format used to elicit expectations of percentage changes, such as return expectations, can significantly influence those reported expectations. We develop a methodology that induces return distributions which serve as a ground truth, and we subsequently vary how we ask participants to report about these distributions. Participants can effectively report differences in means and volatilities irrespective of elicitation method. We find significant differences in elicited levels based on the question format. Directly asking for mean returns positively biases responses. Volatilities based on confidence intervals are not dependable, as responses are largely invariant to the interval size. Asking subjects to report complete distributions yields more accurate responses. We discuss best practices for constructing such questions to obtain meaningful estimates.

The U.S. stock market’s return during the first month of a quarter correlates strongly with returns in future months, but the correlation is negative if the future month is the first month of a quarter, and positive if it is not. These correlations offset on average, consistent with the well-known nearly zero unconditional autocorrelation, yet they are pervasive, present across industries and international markets. The pattern accords with a model in which investors extrapolate announced earnings to predict future earnings, not recognizing that earnings in the first month of a quarter are inherently less predictable. Survey data support the model.