Learning and Risk Premiums in an Arbitrage-free Term Structure Model

We study the evolution of risk premiums on US Treasury bonds from the perspective of a real-time Bayesian learner RA who updates her beliefs using a dynamic term structure model. Learning about the historical dynamics of yields led to substantial variation in RA’s subjective risk premiums. Moreover, she gained substantial forecasting power by conditioning her learning on measures of disagreement among professional forecasters about future yields. This gain was distinct from the (much weaker) forecasting power of macroeconomic information. RA’s views about the pricing distribution of yields remained nearly constant over time. Her learning rule outperformed consensus forecasts of market professionals, particularly following U.S. recessions.