Abstract

We propose a new semiparametric nonlinear di¤usion model based on Reducible Stochastic Differential Equations (RSDEs). The idea of using RSDEs for modelling diffusions was advocated by Bu et al. (2011) who developed two classes of RSDEs, namely the OU-reducible and CIR-reducible SDEs with parametric transformation functions. In this paper, we extend the two classes of models by allowing the transformation functions to be completely unspeci…ed. In contrast to most existing semiparametric models in the literature which typically impose complete parametric structure on either the drift or the diffusion function, in our approach both functions are semiparametric. This means that the shape of neither of the two functions is determined by prior parametric assumptions. A nonparametric estimator of the transformation function is developed, which are then plugged into the likelihood function of the discretely observed process to form a Pseudo-Likelihood function. A Pseudo Maximum Likelihood Estimator (PMLE) can then be obtained for the parameters of the basic processes by maximizing this Pseudo-Likelihood function. The …nite sample properties of the PMLE are examined via Monte Carlo simulations and the quality of the semiparametric transition density estimates is also examined. We also discussed how to extend our framework to cater for SDEs that are reducible to the Aït-Sahalia (1996b) general speci…cation.