The cross-correlation calculated is the biased version, which has lower variance than the unbiased estimator. It reduces towards zero as the lag increases.
If data variances are available then these may be used to weight
the data values, and the expected noise contribution can be
removed from the denominator of the cross-correlation to give an
estimate of the source CCF rather than the data CCF - see e.g.
Weisskopf et al, Ap.J.199, L147.