Time Series Analysis of Non-Gaussian Observations Based on State Space Models from Both Classical and Bayesian Perspectives (Q4943405)

This paper discusses the analysis of non-Gaussian time series using state space models (ssm) from both classical and Bayesian points of view. To start with, in section 2 the linear Gaussian ssm is stated through the observation equation (oe) \(y_t=Z_t \alpha_t+ \varepsilon_t\) and the system equation (se) \(\alpha_t= T_t\alpha_{t-1}+ R_t\eta_t\). Then the nonlinear model is given by the same se and the modified oe \(p(y_t\mid \alpha_1, \dots, \alpha_t, y_1,\dots, y_{t-1})= p(y_t\mid Z_t\alpha_t)\). Two distributional forms are payed special attention to, namely \[ p(y_t\mid \theta_t)= \exp\bigl[y_t' \theta_t- b_t(\theta_t) +c_t(y_t)\bigr] \quad \text{and} \quad y_t= \theta_t+ \varepsilon_t. \] Section 3 gives the basic simulation formulae and in section 4 the Gaussian importance densities are obtained that are needed for simulation by constructing linear Gaussian models which approximate the non-Gaussian model in the neighborhood of the conditional mode of \(\alpha\) given \(y\). Then some further computational methods are discussed. The last section gives real data illustrations. The contributions in the discussion touch all the points of the paper. There is no consensus about the approach but the paper gave rise to diverse remarks concerning this methodology.