The Shtarkov or "Normalized Maximum Likelihood (NML)" code plays a
fundamental role in universal coding and in Rissanen's minimum
description length (MDL) principle for model selection and
prediction. The NML code achieves the minimax regret relative to the
given statistical model. Unfortunately, for many important models -
including models as simple as the normal family - the NML code is
undefined, and the approach has to be modified. We present a new
generalization of NML which provides a generic solution to this
problem: the 'luckiness NML' (related to the 'luckiness principle' in
learning theory) and its special case, the 'conditional
NML'. Luckiness NML can also be applied to non-parametric model
classes such as Gaussian processes. Conditional NML is related to
objective Bayes approaches with improper multidimensional Jeffreys'
prior, and sheds new light on the old question whether it makes sense
to use such a prior.