Estimating Discrete-Time Gaussian Term Structure Models in Canonical Companion Form

File pdf 783KB

This article proposes a convenient parametrization for the popular class of (discrete-time)
essentially-affine term structure models of Duffee (2002), and Ang and Piazzesi (2003). First, I
show that if bond prices are determined by N latent state variables, all their pricing information
must be present in the short end of the term structure, i.e., one can rotate the model to an observationally
equivalent form with exactly N short maturity forward rates as factors. Second, the
risk-neutral transition of the rotated model is conveniently parametrized by N unrestricted real
numbers (contained in a companion matrix), and only one additional parameter is needed to specify
the risk neutral drift. Third, the resulting state-space representation makes it easy to estimate the
model either by the Kalman filter (in one step), or treating N linear combinations of observable
bond prices (or yields) as observable factors. Finally, I interpret some difficulties in fitting the
essentially-affine term structure models to the data, using the standard set of Fama-Bliss discount
bonds as example. The problem is the existence of (spanned) factors that significantly predict term
structure movements but are virtually impossible to detect from the shortest-maturity forward
rates, which is (by the results of this paper) inconsistent with no arbitrage within the discussed
class of models.

Author(s):

Juliusz Radwanski    
HU Berlin
Germany