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.

236 Chapter 7 Multivariate Time Series<br />

Writing ˆγij (h) for the (i, j)-component of ˆƔ(h), i, j 1, 2,..., we estimate the<br />

cross-correlations by<br />

ˆρij (h) ˆγij (h)( ˆγii(0) ˆγjj(0)) −1/2 .<br />

If i j, then ˆρij reduces to the sample autocorrelation function of the ith series.<br />

Derivation of the large-sample properties of ˆγij and ˆρij is quite complicated in<br />

general. Here we shall simply note one result that is of particular importance for<br />

testing the independence of two component series. For details of the proof of this and<br />

related results, see TSTM.<br />

Theorem 7.3.1 Let {Xt} be the bivariate time series whose components are defined by<br />

and<br />

Xt1 <br />

Xt2 <br />

∞<br />

k−∞<br />

∞<br />

k−∞<br />

αkZt−k,1, {Zt1} ∼IID 0,σ 2<br />

1 ,<br />

βkZt−k,2, {Zt2} ∼IID 0,σ 2<br />

2 ,<br />

where the two sequences {Zt1} and {Zt2} are independent, <br />

k |αk|<br />

<br />

< ∞, and<br />

k |βk| < ∞.<br />

Then for all integers h and k with h k, the random variables n1/2 ˆρ12(h)<br />

and n1/2 <br />

ˆρ12(k) are approximately bivariate normal with mean 0, variance<br />

∞<br />

j−∞ ρ11(j)ρ22(j), and covariance ∞ j−∞ ρ11(j)ρ22(j + k − h), for n large.<br />

[For a related result that does not require the independence of the two series {Xt1}<br />

and {Xt2} see Theorem 7.3.2 below.]<br />

Theorem 7.3.1 is useful in testing for correlation between two time series. If one<br />

of the two processes in the theorem is white noise, then it follows at once from the<br />

theorem that ˆρ12(h) is approximately normally distributed with mean 0 and variance<br />

1/n, in which case it is straightforward to test the hypothesis that ρ12(h) 0. However,<br />

if neither process is white noise, then a value of ˆρ12(h) that is large relative to n −1/2 does<br />

not necessarily indicate that ρ12(h) is different from zero. For example, suppose that<br />

{Xt1} and {Xt2} are two independent AR(1) processes with ρ11(h) ρ22(h) .8 |h| .<br />

Then the large-sample variance of ˆρ12(h) is n −1 1 + 2 ∞<br />

k1 (.64)k 4.556n −1 .It<br />

would therefore not be surprising to observe a value of ˆρ12(h) as large as 3n −1/2<br />

even though {Xt1} and {Xt2} are independent. If on the other hand, ρ11(h) .8 |h|<br />

and ρ22(h) (−.8) |h| , then the large-sample variance of ˆρ12(h) is .2195n −1 , and an<br />

observed value of 3n −1/2 for ˆρ12(h) would be very unlikely.

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

Saved successfully!

Ooh no, something went wrong!