Model Averaging using performance and distance measures

In this work we introduce the problem of forecast combination using performance and distance measures for binary outcome. The thesis is focused on model averaging for parametric and non parametric approaches, with a special attention on temporal dependent and independent models. In terms of results, we combine single models using performance measures and we investigate how distance measure based on the Mahalanobis distance can lead to interesting results for model combination. In order to assess the stability and the predictive capability of the models at hand, we employ different cross-validation techniques: Bootstrap cross-validation, 10-fold cross validation and Leave One Out cross-validation. Empirical evidence are give on a real application to predict default probabilities of Small and Medium Enterprises.