Next: Bibliography
Up: Choice of parameters
Previous: Window width
To choose values of the lag
and the group
of indices of
the eigenvectors, we have to follow standard SSA recommendations.
(For an extensive discussion of this problem we refer to
Golyandina, Nekrutkin and Zhigljavsky (2001).) The separability
characteristics (including the
-correlation) play a very
important role here.
If
is not very large, which should be regarded as the most
interesting case in practice, we choose
(recall that
is assumed to be even) and
, where
is such
that the first
components describe well the signal and the
lower
components correspond to noise.
To choose
, visual inspection of the SSA decomposition of the
whole series and some large parts of the series before applying
the change-point detection algorithms is advised. Alternatively,
if the problem is really sequential and a preliminary study of the
time series is not possible, then the recommendation is to use all
the visual SSA tools in the first part of the series to choose
.
If
is too small (underfitting), then we miss a part of the
signal and therefore we can miss a change (the change may occur
in the underestimated components). Alternatively, if
is too
large (overfitting), then we approximate a part of noise together
with the signal and therefore finding a change in signal becomes
more difficult.
Next: Bibliography
Up: Choice of parameters
Previous: Window width