Learning and the Role of Macroeconomic Factors in the Term Structure of Interest Rates

Models of the term structure based on only observable variables have had limited success in explaining movements in longer-term interest rates. A key assumption in much of this literature is that agents know all the parameters describing the model of the economy and that these parameters are fixed for all time. In this paper, we relax both of these assumptions and assume that agents regularly re-estimate the parameters of their models–both those determining the point forecasts and those describing economic volatility–based on incoming data. In this way, we allow for the real-time problem of pricing assets based on the information set available at the time. In addition, we allow for discounting of past data reflecting a concern on the part of agents for structural change in the economy. We find that the learning model with discounting does a much better job at explaining longer-term yields than an equivalent model with constant coefficients estimated over the full sample; in particular, the deviations from the expectations hypothesis are much smaller on average with our learning model. We then estimate the model term premia imposing an affine arbitrage-free structure. We show that incorporating learning improves the in-sample fit and forecasting performance of the model. Learning also implies time variation in the real-time estimates of macroeconomic volatility that is absent in standard macroeconomic models with fixed coefficients. We find that these shifts in estimates of macroeconomic volatility help explain movements in term premia over the past half century. More generally, our analysis highlights the importance of taking into account the information sets of investors in understanding the determinants of bond prices. JEL Codes: D83, D84, G12