Training neural networks for deriving bond rating formulas

Summary form only given. A practical technique is given for extracting simple formulas by first structuring and then training a neural network for bond rating by backpropagation and then reducing the number of needed features to a minimum. A database with 126 patterns was studied. Each pattern has seven features for characterizing bonds in terms of each company´s financial parameters. The feedforward network resulting from applying a parameter reduction technique has only one hidden unit, which is finally dependent on only two input features. The final reduced model network achieves a classification accuracy of 75% in assigning the seven distinct bond ratings. From the reduced network, a compact formula is derived for computing bond ratings. The formula depends on only two variables, which are systematically selected from the seven originally provided for training the neural network