In this paper we evaluate the prediction capability of various neural network models. The models examined in this study differ on the following parameters: data span, learning technique and number of iterations. The neural net prediction capabilities are also compared to results obtained by classical discriminant analysis models. The specific case evaluated is bankruptcy prediction. The common assumption of all bankruptcy prediction models is that fundamental economic factors and the characteristics of a firm are reflected in its financial statements. Therefore, using analytic tools and data from the firm's financial reports, one can evaluate and predict its future financial status. Since the number of bankrupt firms is limited, we used examples (financial statements) from various periods preceding the bankruptcy event. Although the financial statements from the bankruptcy period convey more information, financial statements form distinct periods always improved the models. The prediction capability of the models is improved by using enhanced learning techniques. However, if the enhancement learning technique is too 'strong', the model becomes too specific for the training data set and thus loses its prediction capabilities.
- Linear discriminant analysis
- Neural networks
- Quadratic discriminant analysis