I develop a methodology for extracting "narratives" -- clusters of "stories" similar in meaning -- from unstructured text and apply it to 21,771 news articles from the 2025 Liberation Day episode. Narratives endogenously organize into prediction-oriented clusters, with narratives’ market informativeness convex in the coherence of their predictions. I introduce two dimensions of narrative disagreement: directional disagreement and dispersion. Even controlling for underlying informational content, dispersion and directional disagreement individually increase turnover, but simultaneously high dispersion and directional disagreement have a negative marginal effect on trade. Returns levels are characterized only by average beliefs, while return dynamics are sensitive to narrative disagreement: autocorrelation in returns attenuates in the share of disagreement-attributable trade volume.