Myers ([3]) summarized the
cross-validation statistics that are usually computed and their expected
values:
should be close to zero;
should be small;
, the sample correlation coefficient, should be close to
1;
the normalized estimation error, given by
at each site x, where is the (co)kriging standard
deviation, should be have a mean-square close to 1, i.e. ideally
the estimates should have statistics similar to (but not
necessarily the same as) those of the true values, : that is,
the means, standard deviations, extremes, skewness, etc. correspond
``to a reasonable degree''.