Title:Bayesian estimation of incompletely observed diffusions

Abstract:We present a general framework for Bayesian estimation of incompletely observed multivariate diffusion processes. Observations are assumed to be discrete in time, noisy and incomplete, in the sense that at each observation time a fixed linear combination of the components is observed with noise. If the components are fully observed at discrete times, a data-augmentation approach in which diffusion bridges serve as added auxiliary variables can be used. But in the case of incomplete observations the problem becomes much harder, as diffusion bridges can no longer be sampled independently. We present an adjusted data-augmentation algorithm, in which bridges are simulated conditional on a noisy incomplete observation at an intermediate time as well. We show that the setup of Schauer et al. (2016) and van der Meulen and Schauer (2016) can be extended to simulate such bridges using a generalised version of the guided proposals introduced in these papers.