For decades, macroeconomists have searched for shocks that are plausible drivers of business cycles. A recent advance in this quest has been to explore uncertainty shocks. Researchers use a variety of forecast and volatility data to justify heteroskedastic shocks in a model, which can then generate realistic cyclical fluctuations. But the relevant measure of uncertainty in most models is the conditional variance of a forecast. When agents form such forecasts with state, parameter and model uncertainty, neither forecast dispersion nor innovation volatilities are good proxies for conditional forecast variance. We use observable data to select and estimate a forecasting model and then ask the model to inform us about what uncertainty shocks look like and why they arise.