Calculates either the sum of squares, mean absolute value, or median absolute value
of the set of holdout sample prediction errors, for regression or time series models.
This is only for use as a criterion to be included in an msc.List argument
to TMC, and cannot be used by itself.
This is intended to provide a more accurate estimate of out-of-sample prediction error. The choice of whether to use the sum of squares, median absolute value, or mean absolute value will depend on the particular application, and is analogous to the choice of doing least-squares versus least-median-residual regression. (Note, however, that there is no substantial computational difference between these criteria.)