Learning About the Long Run

Forecasts of professional forecasters are anomalous: they are biased, forecast errors are autocorrelated, and forecast revisions predict forecast errors. Sticky or noisy information models seem like unlikely explanations for these anomalies: professional forecasters pay attention constantly and have precise knowledge of the data in question. We propose that these anomalies arise because professional forecasters don’t know the model that generates the data. We show that Bayesian agents learning about hard-to-learn features of the data generating process (low frequency behavior) can generate all the prominent aggregate anomalies emphasized in the literature. We show this for two applications: professional forecasts of nominal interest rates for the sample period 1980-2019 and CBO forecasts of GDP growth for the sample period 1976- 2019. Our learning model for interest rates also provides an explanation for deviations from the expectations hypothesis of the term structure that does not rely on time-variation in risk premia.