Using data for U.S. and Canada, we find evidence of the time-varying nature of risk premia, which are obtained as difference between long term interest rates and their expected values. We then apply Kalman filtering to extract the conditional variance of term premia prediction errors; our results highlight that this variable is informative beyond term premia and spreads, and it significantly improves upon prediction capability of standard models. In particular, the conditional variance of term premia, reflecting the high volatility of financial markets, anticipates movements in the output growth. Empirical evidence supports the inverse correlation between term premia and business cycle fluctuations. Data suggest that a deterioration of financial markets conditions, as captured by the increased volatility of term premia, anticipates a decline in the output growth. Therefore, term premia conditional volatility has an adverse effect on the economy.

The expectations hypothesis implies that rational investors can predict future changes in interest rates by simply observing the yield spread. According to Mishkin (1990) the expectations theory can also be reformulated in terms of the ability of the spread to predict future inflation. Unfortunately, although appealing, the theory has found little empirical support. Time-varying term premia and changing risk perception have been advocated to rationalize the aforementioned weak empirical evidence. In this work we suggest that the time-varying nature of term premia makes single-equation models inappropriate to analyse the informative content of the term structure. In particular, when the deviations between the expected and the actual spread are large, which occurs in times of soaring term premia volatility, linear models fail to support the expectations theory. Within a threshold model for term premia, we provide evidence that the yield spread contains valuable information to predict future interest rates changes once the risk-averse attitude of economic agents is appropriately considered. Empirical results show that the predictive ability of the yield spread is contingent on the level of uncertainty as captured by the size of monetary policy surprise. JEL classification: C01, C30, E43, G12.