This paper shows how the information that is present in many expectations datasets can be leveraged in a new way in order to extract empirical measures of beliefs about key economic quantities: shocks and their dynamic effects. The information needed is a panel of expectations revisions of one variable across different horizons and over time. The idea is to fit a (time-varying) factor model to the revisions and obtain nonparametric estimates of the latent shocks (the factors) and the associated impulse responses (the loadings) at every point in time. The method relies on weak assumptions and deals with the small-sample nature of these data. An application to consensus inflation expectations reveals a single perceived shock that is highly correlated with inflation surprises and time-varying impulse responses that imply a secular decrease in the perceived persistence of the shock (that is, more “anchored” long-run inflation expectations).