Estimation and Prediction for a Class of Dynamic Nonlinear Statistical Models

A class of dynamic, nonlinear, statistical models is introduced for the analysis of univariate time series. A distinguishing feature of the models is their reliance on only one primary source of randomness: a sequence of independent and identically distributed normal disturbances. It is established that the models are conditionally Gaussian. This fact is used to define a conditional maximum likelihood method of estimation and prediction. A particular member of the class is shown to provide the statistical foundations for the multiplicative Holt-Winters method of forecasting. This knowledge is exploited to provide methods for computing prediction intervals to accompany the more usual point predictions obtained from the Holt-Winters method. The methods of estimation and prediction are evaluated by simulation. They are also illustrated with an application to Canadian retail sales.