Content area
Abstract
The construction industry is one of the major industries in the UK and it has the highest percentage of company failures each year. Systematic investigation of the whole process of company failure prediction and development of a robust and consistent methodology for identifying likely failures of UK construction companies is therefore critical and is the main objective of this study. Data analysis, a novel classifier, a non-failed company selection strategy and a consistent prediction methodology form the four main parts of this thesis. The data analysis was done to investigate the characteristics of the data space spanned by the financial ratios derived from the annual accounts of UK private construction companies and is mainly composed of linearity analysis and overlap measurement between the failed and non-failed companies. The linearity analysis proposed in this study aimed to find the basic relationship between the linearity of the data space and distribution characteristics of the financial ratio.
A method for measuring the degree of overlap between the failed and non-failed companies is proposed for the purpose of revealing the relationship between the overlap degree and the misclassification rate. A homoscedastic model is unable to deal with heteroscedastic data spaces and heteroscedastic classifiers often experience numerical difficulty; the jack-knife, a robust statistic, was therefore introduced to build a novel robust heteroscedastic Parzen window classifier for company failure prediction. Multiple-output models and classification trees were studied so as to avoid inconsistent predictions encountered by multiple discriminant analysis models. It has been found that the methodology of randomly selecting the non-failed companies can have the same, even better, performance compared with the methodology of selecting those non-failed companies whose turnover size, accounting year match with the failed companies. Several case studies are conducted to illustrate the reliability of the novel classifier developed in this study.