Accurate prognosis and prediction of response to therapy are essential for personalized treatment of cancer. Even though many prognostic gene lists and predictors have been proposed, especially for breast cancer, highthroughput "omic" methods have so far not revolutionized clinical practice, and their clinical utility has not been satisfactorily established. Different prognostic gene lists have very few shared genes, the biological meaning of most signatures is unclear, and the published success rates are considered to be overoptimistic. This review examines critically the manner in which prognostic classifiers are derived using machine-learning methods and suggests reasons for the shortcomings and problems listed above. Two approaches that may hold hope for obtaining improved prognosis are presented. Both are based on using existing prior knowledge; one proposes combining molecular "omic" predictors with established clinical ones, and the second infers biologically relevant pathway deregulation scores for each tumor from expression data, and uses this representation to study and stratify individual tumors. Approaches such as the second one are referred to in the physics literature as "phenomenology"; they will, hopefully, play a significant role in future studies of cancer.