In prognostic's field, the lack of knowledge on the behavior of equipments can impede the development of classical dependability analysis, or the building of effective physic-based models. Following that, artificial neural networks (ANNs) appear to be well suited since they can learn from data gathered from equipments. In this paper, an approach combining a Recurrent Radial Basis Function network (RRBF) and an AutoRegressive with eXogenous inputs model (ARX) is proposed in order to perform the prediction step of prognostics: the ARX attempts to correct the error of predictions of the RRBF. Moreover, since performances of an ANN can be closely related to initial parameterization of the network, a criterion is defined to quantify the reproducibility of predictions and thereby a priori estimate the usefulness of neural network structure. The whole aims at improving the prediction step of prognostics, which is critical with respects to real applicative conditions.