Title: Semiparametric prediction intervals in parametric models with non-normal additive error terms
Authors: Gerhard Fechteler - Universität Konstanz (Germany) [presenting]
Abstract: The asymptotic distribution of mean predictions in parametric regression models with additive error term structure is usually well known and often normal. The construction of confidence intervals based on the asymptotic distribution is straight forward. To account for the uncertainty resulting from the error terms, prediction intervals are often more meaningful in applied work. Prediction intervals are commonly constructed for normally distributed error terms in the literature. We propose a simple framework for constructing prediction intervals for non-normal error term distributions. We show that the interval is based on a distribution resulting from the convolution of the distributions of the mean prediction and the error term. The estimation strategy is based on a kernel density estimation of the error term distribution. The implementation is straight forward and applicable to all regression models with known (asymptotic) parameter distribution. We demonstrate the usefulness of the framework via an application to the prediction of house prices.