Recursive estimation and modelling of nonstationary and nonlinear time series

This paper presents a unified approach to nonlinear and nonstationary time-series analysis for a fairly wide class of linear time variable parameter (TVP) or nonlinear systems. The method theory exploits recursive filtering and fixed interval smoothing algorithms to derive TVP linear model approximations to the nonlinear or nonstationary stochastic system, on the basis of data obtained from the system during planned experiments or passive monitoring exercises. This TVP model includes the State Dependent type of Model (SDM) as a special case, and two particular SDM forms, due to Priestly and Young, are discussed in detail. The paper concludes with three practical examples: the first based on the modelling of data from a simulated nonlinear growth equation; the second concerned with the adaptive forecasting and smoothing of the Box-Jenkins Airline Passenger data; and the third providing a critical appraisal of state dependent modelling applied to the famous Sunspot time-series.