Abstract: Dynamic Nelson-Siegel term structure model (DNS hereafter)of interest rate has been extensively studied and widely applied because of its strong ability to fit yield curve. The DNS model has been reformulated as a linear Gaussian state space model with error terms of constant variances both in measurement and transition equations, which is not able to capture the dynamics in conditional covariance of yields. In this article, the conditional heteroscedasticity in the error terms of the DNS model is introduced to improve its fitting performance and forecasting ability. We extend the model by specifying the dynamic models of error term variances in measurement equation as GARCH and that in transition equation as general autoregressive score (GAS) proposed by Creal et al. (2013) using observation-driven method. The decomposition and reparameterization methods are employed to ensure the positiveness of conditional variance matrix. The empirical evidence of considerable increase in within-sample fitting and out-of-sample forecasting goodness for these advances is present in the dynamic Nelson-Siegel model in China bond market.

Key words: term structure model, Nelson-Siegel model, adaptive state space model, GAS model