Summary: The paper present first steps towards linear statistical inference with random fuzzy data. In establishing best linear unbiased estimators (BLUE) or best linear predictors (BLP) it is necessary to consider a suitable notion of expectation and variance for random fuzzy variables. Using \(\alpha\)-cuts of the random fuzzy variables, the mathematical problem can be formulated within the theory of random closed sets (RCS). As a methodological guide we use the Fréchet-approach which leads for a given metric to an associated expectation and variance. In particular, we prove that the well known Aumann expectation is a Fréchet expectation and that the associated variance appears as a reasonable definition for a variance of RCS’s. this associated variance has the advantage that at least special fuzzy number data can formally be handled like Euclidean vectors. Due to the special addition and scalar multiplication these “vectors”, however, fail to build a linear space which leads to somewhat more complicated optimization procedures for BLUE and BLP. This is demonstrated for linear regression and for prediction of stationary processes. In special cases, however, the fuzzified version of the classical BLUE and BLB keep their optimality.