Asymptotically efficient prediction for LAN families

Abstract

In a previous paper (Bosq & Onzon (2012)) we did a first generalization of the concept of asymptotic efficiency for statistical prediction, i.e. for the problems where the unknown quantity to infer is not deterministic but random. However, in some instances, the assumptions we made were not easy to verify. Here we give proofs of similar results based on quite a different set of assumptions. The model is required to be a LAN family, which allows to use the convolution theorem of H\'ajek and Le Cam. The results are applied to the forecasting of a bivariate Ornstein-Uhlenbeck process, for which the assumptions of Bosq & Onzon (2012) are tricky to verify.