The short answer to the original question is that the CAPM is an ex ante model - i.e. it relates the expected return to systematic risk. It is more properly written as E[Ri]=Rf+B(E[Rm]-Rf). We expect returns (including E[Rm]) to be positive, but ex post they may not be. The model is of limited usefulness because it expressses the unobservable expected return E[Ri] in terms of another unobservable, E[Rm]. That is to say, it is not reduced form. Further, there is no practical proxy for E[Rm] because, in theory, it contains all risk assets, not just financial.