Bond rating formulas derived through simplifying a trained neural network
Alvin J. Surkan, Xingren Ying · 1991
The authors report success in reducing a trained multilayered neural network to one that has very few neurons. The simplified network is replaced by a compact analytical formula. In the case of a network trained to rate bonds on the basis of financial data, the final reduced network has a single neuron depending on only two features. In spite of the extreme simplicity of the network, good generalization is observed for the classification of the test patterns which are external to the training set. On the basis of a training set of 100 patterns, a very simple network and formula was derived to put the original full set of 126 patterns of the seven bond parameters into seven output classes with an accuracy of 75%. The derived formula is found to generalize very well even though it has few parameters and is based on only two features.>