Properties of Estimates for some simple ARMA models Petrutza Caragea Fall 2009, Iowa State University P. Caragea (ISU) Stat 551 (Handout 9) Fall 2009 1/7 AR models AR(1): Var(φ̂) ≈ AR(2): 1−φ2 n Var(φ̂1 ) ≈ Var(φ̂2 ) ≈ Corr(φ̂ , φ̂ ) ≈ 1 2 P. Caragea (ISU) 1−φ22 n φ1 1−φ2 Stat 551 (Handout 9) Fall 2009 2/7 MA models MA(1): Var(θ̂) ≈ MA(2): 1−θ2 n Var(θ̂1 ) ≈ Var(θ̂2 ) ≈ P. Caragea (ISU) Corr(θ̂1 , θ̂2 ) ≈ 1−θ22 n θ1 1−θ2 Stat 551 (Handout 9) Fall 2009 3/7 ARMA(1,1) model h ih i2 1−φ2 1+φθ Var( φ̂) ≈ n φ+θ h 2ih i2 1+φθ 1−θ Var( θ̂) ≈ n φ+θ √ (1−φ2 )(1−θ2 ) Corr(φ̂, θ̂) ≈ 1+φθ P. Caragea (ISU) Stat 551 (Handout 9) Fall 2009 4/7 Large-Sample Standard Deviations MA(1) model: Comparison MM vs. MLE AR(1) model with various n and φ φ 0.4 0.7 0.9 50 0.13 0.10 0.06 n 100 0.09 0.07 0.04 200 0.06 0.05 0.03 θ 0.25 0.50 0.75 0.90 SDMM /SDMLE 1.07 1.42 2.66 5.33 since large sample variance for the MoM estimator of θ is Var(θ̂) ≈ P. Caragea (ISU) Stat 551 (Handout 9) 1+θ2 +4θ4 +θ6 +θ8 n(1−θ2 )2 Fall 2009 5/7 Some examples in R. AR(1) case >phi=0.4 >ar1.s=arima.sim(model=list(order=c(1,0,0),ar=phi),n=200) >M1=ar(ar1.s,order.max=1,AIC=F,method=’yw’) # method of moments >M2=ar(ar1.s,order.max=1,AIC=F,method=’ols’) # conditional sum of squares >M3=ar(ar1.s,order.max=1,AIC=F,method=’mle’) # maximum likelihood # The AIC option is set to be False otherwise the function will choose the AR order by minimizing AIC, so that zero order might be chosen. #summarize the information from the above functions: >c(phi,M1$ar,M2$ar,M3$ar) #to compute asymptotic std. dev: >(1-M3$ar*M3$ar) / 200 From the help file for function ar: ar Estimated autoregression coefficients for the fitted model. var.pred The prediction variance: an estimate of the portion of the variance of the time series that is not explained by the autoregressive model. P. Caragea (ISU) Stat 551 (Handout 9) Fall 2009 6/7 Some examples in R. ARMA(1,1) case #ARMA(1,1) >phi=0.6;theta=0.3 >arma11.s=arima.sim(model=list(order=c(1,0,1),ar=phi,ma=theta),n=200) >M2=arima(arma11.s, order=c(1,0,1),method=’CSS’) # conditional sum of squares >M3=arima(arma11.s, order=c(1,0,1),method=’ML’)# maximum likelihood #summarize the three methods >cbind(c(phi,theta),M2$coef[1:2],M3$coef[1:2]) P. Caragea (ISU) Stat 551 (Handout 9) Fall 2009 7/7