4-平稳时间序列分析.ppt

Ex4: MA(1) process 1000 observations of a zero-mean Gaussian MA(1) nobs=1000 theta1=0.2 eps=rnorm(nobs) z=eps[2:1000]+theta1*eps[1:999] length(z) ts.plot(z) plot(density(z)) qqnorm(z) acf(z) pacf(z) Box.test(z,lag=10,type=Ljung) repeat same commands with theta1=0.8 repeat same commands with theta1=0.99 Comments ? How to include a mean ? Ex5：ARMA Simulation ? Modeling? Ex1: Sunspots time series Data analysis: Monthly Jan. 1749 - Dec. 1983 sunspots summary(sunspots) frequency(sunspots) start(sunspots) end(sunspots) Specification acf(sunspots,lag.max=240) pacf(sunspots,lag.max=240) Automatic specification by AIC sunc = sunspots-mean(sunspots) sunc.ar=ar(sunc,order.max=100,aic=T) sunc.ar$aic plot(sunc.ar$aic[2:100]) Where is the min ? p = 28 Ex1: Sunspots time series Estimation methods ar(sunc,order.max=100,aic=T,method=“yw”)$order ar(sunc,order.max=100,aic=T,method=“burg”)$order ar(sunc,order.max=100,aic=T,method=“mle”)$order ar(sunc,order.max=100,aic=T,method=“ols”)$order Yule-Walker is OK Ex1: Sunspots time series Validation res=sunc.ar$resid res=res[29:length(res)] res.std=res/sqrt(var(res)) summary(res.std) plot(res.std) acf(res.std) pacf(res.std) Box.test(res.std) plot(density(res.std))) jarque.bera.test(res.std) I Conclusion ? GWN ? Why ? Ex1: Sunspots time series Forecasting horiz=120 prev=predict(sunc.ar,n.ahead=horiz) prev prev.mean=prev$pred+mean(sunspots) ts.plot(prev.mean,sunspots) Confidence bounds (1 ? = 0.95) prev.se=prev$se bornesup=prev.mean+1.96*prev.se borneinf=prev.mean-1.96*prev.se ts.plot(prev.mean,bornesup,borneinf) Remark : Normality assumption Ex2: Nile river series Data analysis: Annual 1871-1970 Nile plot(Nile) acf(Nile,lag.max=50) pacf(Nile,lag.max=50) Box.test(Nile) In favour of a AR(1) model Automatic AR Specification Nilec=Nile-mean(Nile) Nilec.ar=ar(Nilec,lag.max=10) Nilec.ar Conclusion: p? q? Ex2: Nile river series Automatic ARMA specification arima(Nilec,order=c(1,0,0),i