A common aim of clinical research is the identification of patients at different prognosis, so that future treatment protocols may be tailored to patients' risk profiles. Establishing a prognostic classification is a difficult process, especially in rare cancers: the availability of a growing number of candidate prognostic factors may lead to competitive stratifications, whose validation may not be feasible due to small numbers and the need for a long follow-up. Our goal is to illustrate a strategy to compare stratifications, based on the performance in clinically relevant subgroups of patients. We investigated different statistical measures and recommend a strategy based on the Brier Score, a measure of prediction inaccuracy on individual survival. Results on an infant leukaemia study show that this method is flexible and easily applied with common statistical software. The method, however, does not overcome the problem of lack of validation on external data.