| Summary: | The research uses neural nets as a tool in the investigation of busienss failure prediction and business performance monitoring. The novelty lies in the introduction of models including qualitative factors as well as financial ratios. In addition, an analysis of data gathered from a new survey is offered. To achieve its objectives the research begins by exploring the AI options and then reviews current neural net technology with a view to identifying appropriate technology for the implementation of a classifier for the two areas of failure prediction and performance monitoring. After consideration of the strengths and weaknesses of the options, a multi-layer perceptron, back propagation net is adopted as being unsuitable for this application. In order to verify the validity of the bespoke neural net software it was necessary to employ a two stage strategy. The first step was to confirm that the net, as implemented, retained the expected property of being able to solve problems that were not linearly separable. This was achieved by demonstrating its ability to solve the straightforward XOR problem. To be confident of the net performance it was deemed necessary to replicate the experiments of previous research which used only purely financial inputs to the net. The results confirmed the validity of the new network implementation. Using the intital results as a control, experiments were undertaken to ascertain the effect of reducing the training sample size and to identify minimum sample sizes commensurate with maintaining the effectiveness. The work then further contributes to this research by using traditional stastical methods to provide an empirically derived equation for calculating the minimum number of training patterns required for corporate failure prediction in the context of the experimental sets of variables. The resulting failure prediction model was then used to test for symptoms of bankruptcy in firms currently trading. The thesis then leads on to describing a technique developed in this study for pre-processing qualitative questionnaires, prior to input into a neural model as well as providing a method for predicting values not supplied in incomplete survey responses. A contribution is also made to the area of company performance analysis by using neural techniques and discriminant analysis to show that relationships do exist between certain company variables and business performance, as well as highlighting which of these variables are the most important if an appropriate corporate condition monitoring strategy is to be developed. Lastly, the corporate performance neural network model is enhanced by facilitating the categorisation of a firm into one of several performance bands.
|
|---|
