Within the family of methods named as Singular Spectrum Analysis there exist
several variants of forecasting algorithms for signals corrupted by additive noise.
In this paper the way to estimate the asymptotical accuracy of the recurrent
variant of such a forecast for long series is shown and illustrated. It occurs that almost
all elements of this construction are already studied and published, though some of
them are hard to be implemented in concrete situations. Here all these elements are
brought together, expanded and discussed. Several theoretical and computational
examples are presented.