Learning in Real Time: Theory and Empirical Evidence from the Term Structure of Survey Forecasts

We develop a theoretical framework for understanding how agents form expectations about economic variables with a partially predictable component. Our model incorporates the effect of measurement errors and heterogeneity in individual forecasters' prior beliefs and their information signals and also accounts for agents' learning in real time about past, current and future values of economic variables. We use the model to develop insights into the term structure of forecast errors, and test its implications on a data set comprising survey forecasts of annual GDP growth and inflation with horizons ranging from 1 to 24 months. The model is found to closely match the term structure of forecast errors for consensus beliefs and is able to replicate the cross-sectional dispersion in forecasts of GDP growth but not for inflation - the latter appearing to be too high in the data at short horizons. Our analysis also suggests that agents systematically underestimated the persistent component of GDP growth but overestimated it for inflation during most of the 1990s.