LAUSR

This paper studies an arbitrage-free dynamic term structure model of a Bayesian learner who presumes that dispersion of beliefs is priced in Treasury markets and that measured dispersion has forecasting power for future Treasury yields. The estimated learning rule reveals which aspects of risk and the state of the economy agents learn about and which aspects they (evidently) know from information embedded in the current yield curve. Moreover, we quantify the effect of learning on measured risk premiums in Treasury markets, as well as how an agent pricing with the fitted stochastic discount factor would have adjusted the parameters of her learning rule over business cycles and as the measured degree of disagreement changed over time. Our {\em real-time} model-based learning rule substantially outperforms the consensus forecasts of market professionals over the past twenty-five years, particularly during the years following recessions in the U.S. economy. Moreover, we find that the informativeness of dispersion in beliefs about future yields is distinct from information about the macroeconomy.