Approaches to ARIMA Modeling and Forecasting推荐.pdfVIP

CHAPTER 7 Approaches to ARIMA Modeling and Forecasting he specification and estimation of the conditional mean (i.e., mean esti- Tmate given a certain history of time series observations) is essential in the analysis of a time series. This is, in a first step, typically done in form of an autoregressive (AR) or autoregressive moving average (ARMA) model described in the previous chapter. In case of nonstationarity, we consider autoregressive integrated moving average (ARIMA)(p, d, q) models given by (6.18). To do so, we difference the original level series, possibly nonsta- tionary, until it becomes stationary and model the differenced series in the standard ARMA framework. The original level data can be recovered from the differenced series by integration. There are two basic approaches to provide methods (procedures) for assessing the appropriateness of ARIMA models to describe a given time series. The first approach is attributed to Box and Jenkins.1 In essence, the Box-Jenkins approach involves inspecting the computed sample autocorrelation functions (SACFs) and sample partial autocorre- lation functions (SPACFs) of the time series and comparing them with the theoretical autocorrelation functions (ACFs) and partial autocorre- lation functions (PACFs). Once a good match is observed, respective parameters are computed. The major advantage of this approach lies in its systematic application of steps in model building. The disadvantage is that the visual examination of SACFs and SPACFs is rather subjective. The second approach is to select a set of possible (p, q) combina- tions and estimate the parameters of the corresponding ARMA(p,q) 1 G. E. P. Box and G. M. Jenkins, Time Series Analysis: Forecasting and Cont