04.01.2013 Views

Springer - Read

Springer - Read

Springer - Read

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

5.1 Preliminary Estimation 147<br />

While ˆθ (3)<br />

n is more efficient, ˆθ (2)<br />

n has reasonably good efficiency, except when |θ1| is<br />

close to 1, and can serve as initial value for the nonlinear optimization procedure in<br />

computing the maximum likelihood estimator.<br />

While the method of moments is an effective procedure for fitting autoregressive<br />

models, it does not perform as well for ARMA models with q > 0. From a<br />

computational point of view, it requires as much computing time as the more efficient<br />

estimators based on either the innovations algorithm or the Hannan–Rissanen<br />

procedure and is therefore rarely used except when q 0.<br />

5.1.2 Burg’s Algorithm<br />

The Yule–Walker coefficients ˆφp1,..., ˆφpp are precisely the coefficients of the best<br />

linear predictor of Xp+1 in terms of {Xp,...,X1} under the assumption that the ACF<br />

of {Xt} coincides with the sample ACF at lags 1,...,p.<br />

Burg’s algorithm estimates the PACF {φ11,φ22,...} by successively minimizing<br />

sums of squares of forward and backward one-step prediction errors with respect to the<br />

coefficients φii. Given observations {x1,...,xn} of a stationary zero-mean time series<br />

{Xt} we define ui(t), t i+1,...,n,0≤ i

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!