Title: Tractable estimation and smoothing of highly non-linear dynamic state-space models
Authors: Tom Holden - University of Surrey (United Kingdom) [presenting]
Abstract: A method is presented for tractably estimating large, non-linear dynamic state-space models, such as multi-asset stochastic volatility models, or DSGE models with occasionally binding constraints, based on approximating their likelihood. The method approximates the distribution of unknown states by an extended skew-t distribution, allowing it to track their first four moments without sacrificing the computational and econometric advantages of a parametric approach. We show that the extended skew-t distribution maintains the properties of the Gaussian distribution that gave tractability to the Cubature Kalman Filter (CKF). Our method extends the CKF by introducing alternative cubature procedures, to further improve the tracking of non-linearities, an augmented-state representation, reducing integration requirements, and dynamic state space reduction, to ensure that it can handle the large state spaces generated, for example, by pruned perturbation solutions to medium-scale DSGE models. We also show how a modified objective can deliver efficient, consistent estimates despite the approximations inherent in our method.