5 - Non-parametric methods for credit risk analysis: Neural networks and recursive partitioning techniques

Summary

Introduction

In all credit analysis problems, a common factor is uncertainty about the continuity of the business being analysed. The importance of business continuity in credit analysis is reflected in the focus, by both academics and practitioners, on constructing models that seek to predict business continuity outcomes (failure or distress). There are two types of modelling exercise that can be useful to decision makers. The first are models that generate the probability of default, an important input to expected loss calculations. The second are classification models, which are used in credit-granting decisions. In this chapter we will look at two non-parametric approaches, neural networks for the generation of default probabilities and classification and recursive partitioning for classification. Each method and its implementation will be presented along with a numeric example.

There is an extensive literature that documents problems in empirical default prediction see Zmijewski (1984), Lennox (1999) or Grice and Dugan (2001). One of the earliest issues was the distributional assumptions that underlie parametric methods, particularly in relation to multiple disriminant analysis (MDA) models. There have been a number of attempts to overcome the problem, either by selecting a parametric method with fewer distributional assumptions or by moving to a non-parametric method. The logistic regression approach of Ohlson (1980) and the general hazard function formulation of Shumway (2001) are examples of the first approach.