This paper documents large micro-heterogeneity and forecasting skill in the cross-section of survey based bond risk premia. We reject informationally constrained rational expectations but show a learning model distorted by sentiment is consistent with the data. Aggregating, we propose a belief...