On the Expectations Hypothesis in US Term Structure

We extend the vector autoregression (VAR) based expectations hypothesis (EH) test of term structure, considered in Bekaert & Hodrick (2001), B&H thereafter, using recent developments in bootstrap literature. Modifications include the use of wild bootstrap to allow for conditional heteroskedasticity in the VAR residuals without imposing strict parameterization, while keeping their contemporaneous correlation, endogeneous model selection procedure in the bootstrap replications to reflect true uncertainty and the stationarity correction designed to prevent finite- sample bias adjusted VAR parameters from becoming explosive. Since Lagrange Miltiplier, Wald and Distance Metric test statistics used in this study are all asymptotically pivotal we estimate their finite sample distributions using a computer simulation, rather than relying on the approximation provided by the first order asymptotic theory. When the modified B&H methodology is applied to extensive US zero coupon term structure data ranging from 1 month to 10 years we find less rejections for the theory in a sub-sample of Jan 1982- Dec 2003 than in Jan 1952- Dec 1978, and when it is rejected it occurs at the very short and long ends of the maturity spectrum. It is also relieving to note that this inference seems to be robust to both AIC and SIC model selection methods. In terms of the conclusions made about the validity of the EH of term structure, the main difference between this study and its counterpart of Sarno, Thornton & Valente (2006), which uses the original B&H methodology, is that we reject the theory less often than they do. This is probably as one would expect, since we test the EH theory of term structure only as opposed to Sarno et al (2006) who, in effect, are testing a joint null hypothesis of the conditional homoskedasticity in the residuals and exogenous lag length of the VAR along with the EH.