Robust prediction limits based on M-estimators

In this paper we discuss a robust solution to the problem of prediction. Following Barndorff-Nielsen and Cox (1996) and Vidoni (1998), we propose improved prediction limits based on M-estimators instead of maximum likelihood estimators. To compute these robust prediction limits, the expressions of the bias and variance of an M-estimator are required. Here a general asymptotic approximation for the bias of an M-estimator is derived. Moreover, by means of comparative studies in the context of affine transformation models, we show that the proposed robust procedure for prediction behaves in a similar manner to the classical one when the model is correctly specified, but it is designed to be stable in a neighborhood of the model.