Thursday 12 Mar 2015: Statistical Science Seminar: Reliability of ensemble forecasts revisited -- exact and asymptotic tests
Jochen Broecker - University of Reading
An important notion in the context of ensemble forecasts (e.g. for weather or climate) is forecast reliability. Operational definitions of reliability basically state that the verification should be statistically indistinguishable from the ensemble members. The main deficit with these definitions is that no statement is made about the joint distribution of a series of verification--ensemble pairs. This precludes a precise analysis of reliability tests (both existing and to be devised).
In this talk, a new notion of reliability, called strong reliability, will be presented which alleviates this problem. Strong reliability is a natural extension of existing notions of reliability yet it provides a basis for rigorous treatment of reliability tests. Both existing and new reliability tests will be discussed in this light.