This paper evaluates methods that employ Kalman Filter to estimate
Diebold and Li (2006) extensions in a state-space representation, applying
the Nelson and Siegel (1987) function as measure equation and different
specifications for the transition equation that determines level, slope and
curvature dynamics. The models that were analyzed have the following
structures in transition equation: (1) AR(1) specification, employing a
diagonal covariance matrix for the residuals; (2) VAR(1) specification,
employing a covariance matrix calculated with Cholesky decomposition; (3)
VAR(1) extension, inserting variables related to the Covered Interest Rate
Parity (CIRP); (4) VAR(1) extension, including stochastic volatility
components. The major findings of this paper were: (1) evaluating the latent
variables dynamics, the curvature was the factor that fitted better to the
stochastic volatility component; (2) in a broad sense, even though the
simplest VAR(1) model was the one that provided the best out-of-sample
performance in the most part of maturities and forecasting horizons, the
extension inserting variables related to the CIRP was able to overcome the
former specification in some of these simulations.