Choosing and Combining the Right Trees From an Ensemble

In this paper, we introduce a novel ensemble approach in the spirit of model clustering and combination. After generating a sample of bootstrap-based trees and identified their main clusters, the corresponding medoid trees are next combined by growing a tree on trees, namely a recursive partition using as covariates the predictions of the medoid trees. The resulting Final Classification Tree is then expected to outperform single trees by construction, being realized with the objective to maximize the forecasting power of single trees. Using data on 472 defaults and 471 non-defaults for non-financial corporates in Italy during the period 2007-13, the research experiment proves that our approach significantly outperforms all the representative trees also ranking among the best 1% distribution of the bootstrap-based trees in terms of performance diagnostics.