Neural networks for bond rating improved by multiple hidden layers

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