Excellent article.
One note though: the similarity prediction gives indeed the best results (also in my model GT data mining) but it has one serious shortcoming, which is mentioned in the article, the results remain a black box, they cannot be rationalized and therefore are very hard to implement in management situations.

Managements usually seek to be involved in the subjects of prediction, they need to prove that the decisions taken are better than the alternatives and well reasoned with the people that are part of them.

The ideal is to get both, 'exact prediction' and the list of control factors that generate the predicted result.