Mis-parametrization subsets for a penalized least squares model selection

Xavier Guyon (1) , Cécile Hardouin (1, 2)

1 SAMM - Statistique, Analyse et Modélisation Multidisciplinaire (SAmos-Marin Mersenne)
2 MODAL'X - Modélisation aléatoire de Paris X

When identifying a model by a penalized minimum contrast procedure, we give a description of the over and under fitting parametrization subsets for a least squares contrast. This allows to determine an accurate sequence of penalization rates ensuring good identification. We present applications for the identification of the covariance for a general time series, and for the variogram identification of a geostatistical model.

Mots clés: least squares contrast; penalized contrast; model selection; misfitting; AIC; BIC; mixture models; geostatistics.; geostatistics

Domaines: Statistiques [math.ST]; Théorie [stat.TH]

Xavier Guyon, Cécile Hardouin. Mis-parametrization subsets for a penalized least squares model selection. 2011. ⟨hal-00648151⟩