Score Driven Dynamic Nelson–Siegel Term Structure Model Of Interest Rates And It’s Application

Interest rate is one of the most important variable in money and financial markets,and the term structure or yield curve of interest rate,which unifies interest rate and term,has become the focus of academic researchers,policy makers and market participants because it contains more abundant information.With the improvement of China’s interest rate marketization and the prosperity of the bond market,it is of great practical significance to build an interest rate term structure model that conforms to the reality of China’s interest rate market,accurately predicts the future interest rate trend and effectively extracts the market expected information.Among the interest rate term structure models,dynamic Nelson-Siegel(DNS)interest rate term structure models have been widely used because of their simplicity,easy expansion and good fitting and prediction on the real interest rate.Aiming at the defects of the traditional DNS interest rate term structure model(Diebold&Li,2006;Diebold et al.,2006)that there is a deviation between the assumption of normal distribution and conditional homoskedasticity and the actual situation,this paper makes two generalizations.Firstly,using the setting method of generalized autoregressive score model(Creal et al.,2013),the error term of DNS model is set as t-distribution with stronger ability to describe thick tailed distribution we construct GAS-t-DNS model.Secondly,this paper extends the traditional DNS model into GAS-TVV-DNS model with time-varying variance within the framework of state space model by using matrix decomposition and re-parameterization.The two extended models are applied to the real yield data of Chinese government bonds.The results show that the two extended models have better in sample fitting and out of sample prediction performance than the traditional DNS model.The traditional DNS model usually sets the loading factor as a constant inadvance or static parameter to be estimated,and rarely considers the time-varying of the loading factor,λ.In this paper,based on the score driven time-varying parameter modeling method,the loading factor,λ,is time-varying under the linear Gaussian state space framework,and the GAS-λ-DNS model is constructed.The results show that the loading factor λ shows strong volatility and is closely related to the economic cycle.When the time-varying loading factor lambda is used to predict China’s economic growth rate,the results show that the time-varying loading factor λt has additional incremental information compared with the traditional macro prediction factors.The introduction of loading factor λt can effectively improve the prediction accuracy of the model.