Abstract
Nonparametric techniques, such as kernel estimators and k-Nearest Neighbor (k-NN), are suitable for prediction techniques in large-scale data, since they directly hold the data structure for prediction. They can be robust techniques especially against changes of the data structure. However, the optimal bandwidth selection problem must be solved to enable high performance. Vieu (1991) proposed the optimal bandwidth selection technique for univariate data which is based on the minimization of a local cross-validation criterion. It is capable to be usefully applied for prediction problems. In this paper we expand the technique for multivariate data and show some numerical examples to clarify its validity.
