On calibration of kullback-leibler divergence via prediction

In this paper we evaluate mean Kullback-Leibler divergence via predicting densities arising from various prediction methods applied to the multivariate single-sample normal model. We demonstrate that the degrees of freedom which index Geisser-Cornfield predictive densities are helpful in divergence calibration. Alternative calibrations are derived based on sample size considerations. An application of each method to univariate prediction from the gamma model is provided. Comparisons are made with a probability-based calibration method.