Predicting company failure-a comparison between neural networks and established statistical techniques by applying the McNemar test

The behaviour of a given system may be forecast using two general methodologies. The first depends upon knowledge of the laws that govern a particular phenomenon. When this knowledge is expressed in terms of a precise set of equations, which, in principle can be solved, then, providing that the initial conditions are specified, the future behaviour of the system may be predicted. However, in cases of systems belonging to behavioural science and economics, for example, the rules governing the behaviour of the system are not readily available. A second, less powerful method, involves the discovery of empirical regularities in observations of the system. As emphasised by Refenes, Azema-Barac, Chen and Karoussos (1993), such regularities are often masked by noise, whilst phenomena that seem random, without apparent periodicities, remain recurrent in a generic sense. As with any technology that is readily available, those companies that are using neural networks successfully are probably remaining silent so as to maintain their competitive advantage. As for the less than silent ones, it is doubtful whether they have discovered the advantages that neural networks may offer. This research, using a backpropagation neural network methodology, proposes to establish whether using neural networks to predict company failure is more successful than using established methodologies