Modeling corporate failure prediction using static and dynamic approaches: neural networks, Bayesian networks, duration and dichotomous models.

Summary The objective of this thesis is to study different methods of predicting business failure in both static and dynamic approaches. More precisely, in the static approach, we have resorted to the methods of selection of the discriminating variables by using neural networks. The first one, based on the HVS criterion, entitled HVS-AUC, allowed us i) to build a more parsimonious model compared to the ADL . ii) to identify a set of stable variables that are both non-cyclical and with a high explanatory power. In contrast, the second technique is based on the forward procedure or, more precisely, on forward-AUC. This method produces results comparable to the LDA but with fewer explanatory variables. We also used Bayesian network structure learning methods to try to improve the classification performance of the companies. We mobilized a technique called "Max-Min Hill-Climbing" or MMHC. We analyzed the classification performance of a combined algorithm between MMHC and the basic model of a naive Bayesian network (BN). This new method has been named BN-MMHC (Naive Bayes augmented by MMHC). In the second dynamic approach, we put more emphasis on factors that cannot be measured a priori and on explanatory factors that cannot be understood in a static framework. In the first part, we mobilized macroeconomic variables to better estimate the risk of default. In the second part, we used an alternative model that allows us to correctly assess the shocks that firms may experience over time. In this way, we have evaluated the effect of the propagation of these shocks.