CREDIT RISK MODELING FOR COMPANIES DEFAULT PREDICTION USING NEURAL NETWORKS

The paper assesses the business default risk on a cross-national sample of 3000 companies applying for credit to an international bank operating in Romania. The structure of the sample replicates the structure of the general population of companies in Romania. Based on their past credit history, we have distributed the companies in seven classes plus the default, using and adapting the Standard & Poor's categories: AAA (1020 companies, 34%) - no risk; AA (279 companies, 9.3%) - minimal risk; A (906 companies, 30.2%) - low risk; BBB (201 companies, 6.7%) - moderate risk; BB (123 companies, 4.1%) - acceptable risk; B (111 companies, 3.7%) - high risk; C (105 companies, 3.5%) - very high risk and D (255 companies, 8.5%) - default. We have then, estimated the one-step transitions probability for downgrading for one year, based on the present category, loan amount, size of company and sector of activity. Thus, although the approach is bottom-up and unconditioned, focusing on the companies, we have included the economic context, taking into account the type of company and the economic sector. We have performed the estimations first using logit regression, and then ANN (Artificial Neural Networks), and compared the results with Standard & Poor's transition matrix for 2010. The results were compared in terms of predictive power, and arguments were given for choosing an ANN design.