Title: A rating model simulation for risk analysis
Author: Greta Falavigna
Address: Ceris-CNR, via Real Collegio 30, Moncaliere 10024, Turin, Italy; University of Eastern Piedmont 'Amedeo Avogadro', Novara, Italy
Journal: Int. J. of Business Performance Management, 2008 Vol.10, No.2/3, pp.269 - 299
Abstract: This study analyses the situation of a bank that wants to create an Internal Rating System (IRB). A credit institute can decide to simulate rating judgements from an external rating agency, like Standard and Poor's or Moody's or Fitch Rating. This research compares different frameworks of neural networks, hybrid neuro-fuzzy model and logit/probit model, used to simulate the rating of an external agency. Initially, the models are divided into eight rating classes but the mean percentage error is big. Hence, a two-stage hybrid neuro-fuzzy framework is built, in which the model correctly distinguishes the firms into three macroclasses and then, for each macroclass, a hybrid model divides the firms into eight different classes. This two-stage framework provides good results.
Keywords: feed-forward neural networks; radial basis function; RBF; generalised regression neural network; Grnn; probabilistic neural network; Pnn; multinomial logit; probit; internal rating systems; default risk; complex system; risk assessment; neuro-fuzzy models; hybrid modelling; artifical neural networks; ANNs; insolvency.