4- Nonstationary- Models-and- Regression- In-this-chapter-we-课件.ppt

ARMA, F(Bs)Yt = Q(Bs)Zt, and the nonseasonal component via the ARMA, f(B)Yt = q(B)Zt. These two are then combined multiplicatively as in the definition. The preliminary differencing on Xt to produce Yt, will take care of any seasonal nonstationarity that may occur, e.g. when the process is nearly periodic in the season. SARIMA Modeling Guidelines: With knowledge of s, select appropriate values of d and D in order to make Yt=(1-B)d(1-Bs)DXt appear stationary. (D is rarely more than 1.) Choose P Q so that , h=1,2,…, is compatible with the ACF of an ARMA(P,Q). (P Q typically less than 3.) Choose p q so that is compatible with the ACF of an ARMA(p,q). Choice from among the competing models should be based on AICC and goodness of fit tests. * 。 A more direct approach/alternative to modeling the differenced series {Yt}, is to simply fit a subset ARMA to it without making use of the SARIMA multiplicative structure. The forecasting of SARIMA processes is completely analogous to that of ARIMA’s. Ex: (DEATHS.TSM) Form Yt=(1-B)(1-B12)Xt to obtain a stationary-looking series (s=12, d=D=1). The values suggest an MA(1) (or AR(1)) for the between-year model i.e. P=0, Q=1. Inspection of suggests also an MA(1) (or AR(1)) for the between-month model i.e. p=0, q=1. Our (mean-corrected) proposed model for Yt is therefore Yt = (1 + q1B)(1 + Q1B12) Zt. Based on we make the initial guesses: q1 =- 0.3, Q1=-0.3. This means that our preliminary model is the MA(13): * 。 Yt = (1 - 0.3B)(1 - 0.3B12) Zt = Zt - 0.3Zt-1 - 0.3Zt-12 + 0.09Zt-13. (Preliminary estimation algorithms don’t allow subset models.) Now choose “constrain optimization” in the MLE window, and select 1 in the “specify multiplicative relations” box. Enter 1, 12, 13 to indicate that q1 ? q12 = q13. Final model has AICC=855.5, and {Zt} ? WN(0,94251): Yt = 28.83 + Zt - 0.479Zt-1 - 0.591Zt-12 + 0