ABSTRACT
'Recursive partitioning' is an iterative technique which classifies sequentially 'm' terms of a series of independent variables. This is done by a computer program which separates cases according to the lowest misclassification cost. A natural evolution of the recursive partitioning technique is to develop 'artificial intelligence' models on a computer which discriminate between failed and non-failed concerns. The Neural networking (NN) model correctly identified all failed and non-failed companies in the training sample, compared to a successful classification rate of only 86.8 per cent for a benchmark discriminant model. 'NN' models take the iterative approach one step further. Various NN models were then developed to see how they performed in relation to the multiple discriminant analysis models. One, two and three layer NN models were devised with different numbers of hidden nodes and with different learning times. The topology of the models comprised an input layer, a hidden layer, and an output layer.
