In many data gathering contexts, so little physical modeling information is available that a Statistician might default to a predictive optimality criterion.  Even in such complicated settings, there may be general principles that can be invoked to guide the statistical modeling procedure.

In this talk three heuristic candidate statistical principles will be stated and explored in computational settings.  They are: 1) A `principle of matching' between the model and the function space in a regression context.  2) A generalized variance bias analysis to account for model uncertainty and mis-specification. 3) A data summarization principle relating the number of statistics sample size and cluster number in a short fat data context.

This is very much work in progress, so comments and discussion will be truly appreciated.