The Recursive Partitioning Algorithm (RPA): A Nonparametric Classification System

Abstract

Leo Breiman (Breiman et al., 1984, 1998) was a statistician who was fond of practical applications, and this led him to develop several original studies. Based on the work begun by Friedman (1977), he developed a very accurate classification system, without the need for statistical assumptions, since it is a nonparametric methodology. The aim of this study is to present the work of Breiman known as the Recursive Partitioning Algorithm. The RPA will be introduced as a nonparametric approach to credit analysis, allowing for the incorporation of the costs of misclassifications. Several studies, such as Novak and LaDue (1999) and Marais, Patais and Wolfson (1984), have shown its applicability in the analysis and granting of credit. A long road has been traveled from the early work of Friedman (1977) to the CART model developed by Steinberg and Golovnya (2006). This paper – apart from presenting the fundamentals and possibilities for use of the RPA – seeks to show the effectiveness of the results attained through a comparison with a parametric model, the Discriminant Analysis, considered the most traditional and classical method of analysis. The results show the RPA to be a superior technique, as well as a technique of easy intuition by analysts. The conclusion of the paper confirms that the RPA system – little known and discussed by academics and market professionals – is a powerful classificatory tool, with the advantage of being nonparametric.