Automatic prediction and model selection

Abstract

This dissertation is about Automatic Model building and Prediction procedures that are useful to approximate and forecast the expected conditional mean of a stationary target variable. We review the theoretical foundations of model selection and compare the out-of-sample predictive ability of different automatic selection procedures, focusing especially on the the RETINA method proposed by P´erez-Amaral, Gallo & White (2003). A new software implementation of RETINA called RETINA Winpack is proposed. This software piece is designed for immediate use by non-specialist applied researchers. As an important advantage over the original RETINA implementation,it handles extreme observations and allows for distinctive treatment of categorical inputs. Using RETINA Winpack, we present an empirical application to Telecommunications demand using firm-level data. RETINA Winpack is proven to be useful for model specification search among hundred of candidate inputs and for finding suitable approximations that behave well out-of-sample in comparison with alternative linear baseline models. With the aim of increasing the flexibility of the RETINA method in order to deal with non-linearities in the target variable, a new method called RETINET is presented. It generalizes RETINA by expanding the functional approximating capabilities in a way which is similar to Artificial Neural Networks (ANN), by avoiding some of the difficulties related to their practical implementation. As an advantage over traditional ANN, RETINET’s specifications retain, to some extent, analytical interpretability. Based on two different simulation examples the method provides favorable evidence with respect to the out-of-sample forecasting ability provided by both simpler and/or more complex modeling alternatives. RETINET balances between a) Flexibility b) Parsimony c) Reverse engineering ability, and d) Computational speed. The proposed method is inspired by a Specific to General philosophy, going from the simple to the sophisticatedly simple, avoiding unnecessary complexity.