Estimation of Dynamic Term Structure Models

TLDR

The study examines finite‑sample properties of standard estimation techniques for modern term‑structure models. The authors find that maximum likelihood is accurate for simple models but biased with flexible dynamics, Efficient Method of Moments performs poorly in small samples despite asymptotic efficiency, and the linearized Kalman filter offers a tractable, reasonably accurate alternative when maximum likelihood is impractical.

Abstract

We study the finite-sample properties of some of the standard techniques used to estimate modern term structure models. For sample sizes and models similar to those used in most empirical work, we reach three surprising conclusions. First, while maximum likelihood works well for simple models, it produces strongly biased parameter estimates when the model includes a flexible specification of the dynamics of interest rate risk. Second, despite having the same asymptotic efficiency as maximum likelihood, the small-sample performance of Efficient Method of Moments (a commonly used method for estimating complicated models) is unacceptable even in the simplest term structure settings. Third, the linearized Kalman filter is a tractable and reasonably accurate estimation technique, which we recommend in settings where maximum likelihood is impractical.