Predicting the unpredictable (Q1187225)

A difficulty for existing theories of predictive inferences involves the question of what to do when unsuspected events occur. The standard approaches to such inferences, due to Laplace, Johnson, de Finetti and Carnap, adopt a model in which there is a prespecified number of types (attributes) which may be observed. But, as a matter of fact, this is often not the case. A model proposed by De Morgan in the last Century can overcome this difficulty, but he did not advance any justification for his solution. In the present paper a general approach to this problem is suggested. From this emerges that the old solution has a very deep justification. The key idea of the new approach is that although exchangeable sequences are the right objects to consider when the possible outcome-types are known in advance, exchangeable partitions are the right objects to consider when they are not. The resulting theory has parallel basic elements as the classical one. That is: a representation theorem, a distinguished class of partitions and a fundamental theorem (a rule of succession). In the reviewer's opinion this is a very interesting paper both from a statistical and philosophical point of view. It is also worth noting the importance of foundational studies in the advancement of statistical research.