1. Introduction
Forecasting future interest rates has always been a major concern of both economists and
policymakers. Understanding the dynamics of interest rates is important for financial economists,
because, for instance, the price of derivative securities depends on market’s yields; it is equally
essential for macroeconomists, as long as aggregate-spending decisions depend on long term
interest rates, while the opportunity cost of holding money is represented by short term rates. In
addition, understanding the relationship between short and long term rates is relevant for
policymakers. Although the monetary authority might be interested in influencing long term interest
rates, monetary policy actions can only affect the short end of the term structure; moreover, the
Treasury can eventually perform active debt management, since the maturity structure of public
debt affects the government budget.
In this paper we attempt to rationalize the empirical failure of the expectations hypothesis by
exploiting the potentiality offered by non linear modelling. In particular, this study investigates
whether threshold effects are relevant in the empirical analysis of the expectations theory. Working
with U.S. post-war data in a multiple regime framework we examine the informative content of the
U.S. term structure. Evidence suggests that non linearity can account for the empirical failure of the
expectations hypothesis; in particular, we find that the predictive ability of the yield curve’s slope to
anticipate future movements in short rates is contingent on both the level and the volatility of term
premia. Threshold estimates indicate that the yield spread returns an accurate prediction of future
short term rates when the absolute value of the term premium is low.
The interest in non linear models in general, and in threshold models in particular, is motivated by
the fact that the expectations hypothesis (henceforth EH) has been usually rejected by linear
models. The empirical failure of the EH in single-equation models has been often attributed to the
presence of a time-varying term premium (Mankiw and Miron, 1986; Fama, 1986; Cook and Hahn,
1989; Lee, 1995; Tzavalis and Wickens, 1997). In addition, McCallum (1994, 2006) has shown the
intrinsic inability of linear models to corroborate the EH simply considering a time-varying first-
order autoregressive term premium coupled with a monetary policy rule that allows for interest rate
smoothing. Hence, we believe that the time-varying nature of the term premium might be
considered a significant source of asymmetry in the empirical analysis of the expectations theory. In
this vein, this contribution extends and complements the strand of literature pioneered by Campbell
and Shiller (1991). Although the slope of the term structure gives a forecast in the right direction of
long term changes in short rates the predictive ability of linear models is usually modest. Therefore,
since linear models can be viewed as constrained non linear models, detection of non linearity can
improve the ability of the spread to anticipate the evolution of future interest rates. Results highlight