On models and methods for Bayesian time series analysis (Q1069650)

A new approach to the applied econometric problems of adjusting and forecasting univariate time series with component models is described. The models used differ from the familiar ARIMA class by having components derived from fractionally differenced Gaussian processes and by allowing for simultaneous estimation of deterministic along with random components. Methodology proposed for fitting these models is based on Bayesian principles, with particular emphasis placed on assessing sensitivity of conclusions to model assumptions. The methods require extensive likelihood-related computations, using frequency domain representations to produce a range of new graphical diagnostic displays. They are illustrated with an extended example.