Predictive efficiency for simple non-linear models (Q1118295)

This paper demonstrates the use of exact predictive likelihood functions for simple non-linear models. A measure of predictive efficiency based on the concept of expected information loss is introduced as a way of comparing alternative prediction functions. It is shown that the predictive likelihood function minimizes expected information loss over a wide class of potential prediction functions. Some Monte Carlo experiments illustrate the performance of alternative prediction functions in settings where prediction is difficult.