Predictive performance for the detection of underfitting in density estimation

A methodology for model selection and validation is presented. The technique is applied to probability density estimation using Radial Basis Function (RBF) Neural Network. A model is assumed to have a good-fit, if it is able to predict the data. We present a procedure to test if the prediction of the model is calibrated, i.e. if the predicted data frequencies match the empirical data frequencies. Some experimental results show the benefits of such an approach.