Discretised Non-Linear Filtering for Dynamic Latent Variable Models: with Application to Stochastic Volatility

Filtering techniques are often applied to the estimation of dynamic latent variable models. However, these techniques are often based on a set assumptions which restrict models to be specified in a linear state-space form. Numerical filtering techniques have been propsed that avoid invoking such restrictive assumptions, thus permitting a wider class of latent variable models to be considered. This paper proposes an accurate yet computationally efficient numerical filtering algorithm (based on a discretisation of the state space) for estimating the general class of dynamic latent variable models. The empirical performance of this algorithm is considered within the context of the stochastic volatility model. It is found that the proposed algorithm outperforms a number of accepted procedures in terms of volatility forecasti