Section: New Results

Cross validation algorithms

Participant : Alain Célisse.

The performance of Cross-validation (CV) algorithms are assessed for estimating the risk as well as for model selection. Whereas optimality of leave-one-out (LOO) cross-validation is proved for risk estimation, it is no longer the case for model selection. In the latter setup, conditions are derived that lead to optimality for leave-$p$-cross-validation (LPO) when $p$ is larger than 1. See for details [47] .