Analysing time series for forecasting (a personal view)

This paper discusses the author’s views on orthodox time-domain modelling practice, together with suggested modifications and extensions to the now well-established Box-Jenkins methodology for analysing and forecasting time series. We concentrate on improving interpretation of the serial correlation structure for the purpose of enhancing model identification. For instance, by more sensitive discrimination between stationary and nonstationary situations, we may decrease expected forecast error when dealing with long-memory processes.