On inference for partially observed non-linear diffusion models
using the Metropolis-Hastings algorithm
Osnat Stramer
Department of Statistics and Actuarial Science
University of Iowa
We develop a new approach for the inference of a broad
class of non-linear continuous time models, when the data are observed
at discrete time points. We employ a Bayesian approach for
model estimation based on MCMC methods. We use the
Hastings-Metropolis algorithm with collection of Brownian bridges
and linear diffusion bridges as candidates for the
independent sampler to obtain
data augmentation algorithms for the missing paths between any two
observations. We thus obtain algorithms which are independent of the
sample intervals. Our approaches are illustrated with examples
involving simulated data, and are applied to two data sets: the
US interest rates and
the IBM closing stock prices respectively.