The failure of corporate failure models to classify and predict : aspects and refinements
Degree GrantorUniversity of Canterbury
Degree NameMaster of Commerce
Much has been written about the use of multiple discriminant analysis in corporate distress classification and forecasting. Classification and prediction models are notoriously difficult to establish in such a way that they will stand the ultimate test of time. Many articles severely criticise the use of the technique yet there are aspects which may improve our ability to develop satisfactory models. We are probably yet a long way off from being able to do so with any great degree of satisfaction, yet it behoves us to try to develop models that do justice to the assumptions and the theory. This thesis explores several important aspects of the model-building process and concludes that some of the more conventional criticisms of the models developed so far are less important than claimed. It suggests that more critical than the failure to meet the conditions of multivariate normality, the equality of the variance-covariance matrices, and the use of a priori probabilities are the need for: a satisfactory model specification that can be theoretically justified, the strict use of random sampling, the efficient use of sample data, the search for stable mean vectors which are significantly different from each other, and ex ante validation. If these requirements are met then the MDA technique is robust enough to cope with breaches of the assumptions.