Neural networks for bond rating improved by multiple hidden layers

The problem of how best to configure and train a network to rate bonds on the basis of financial parameters which characterize corporations is addressed. 126 bond patterns, each with seven financial ratios, were collected. A spectrum of layered neural networks with single or multiple hidden layers was found to easily outperform a bond rating model based on multivariate discriminant analysis, at least when the task was presented as one of separating two noncontiguous classes. These experiments suggest that networks trained with two hidden layers outperform a network having only one hidden layer containing a comparable number of processing elements. Significant advantages arise even when the layers are ordered so that a smaller number of neurons receive their inputs directly from the inputs. This suggests that, for these bond rating data. the problem of mapping from seven features to two classes may have an inherent dimensionality of five or lower. The mapping of the five elements' (or neurons') output of the first layer down to two layers can be performed first. Then there is still an advantage from the use of a subsequent hidden layer to provide another representation of useful classification information