Predicting corporate failure through a combination of intelligent techniques

2017-02-23T04:35:17Z (GMT) by Gunnersen, Sverre Edvard
Corporate failure is one of the most popular prediction problems because early identification of at-risk companies presents such a clear economic benefit to creditors, investors and society as a whole. Throughout the years statistically based classification systems, intelligent systems such as Neural Networks with its many variants, and newer techniques such as Genetic Programming have been applied to this problem. Indeed when a new variation or technique is proposed, the prediction of corporate failure is often one of the first test domains for the new methodology. Likewise, the cause of corporate failure is a topic that has received much academic and literary attention, including case studies investigating the trajectories that failing companies take or post hoc qualitative analysis as to whether certain fundamental causes such as one-man-rule can be attributed to the subsequent collapse of a company. However, throughout the history of this topic a number of challenges emerge that remain unaddressed within the literature. The first challenge is that while many papers outlining new classification techniques compare results with another popular classification system as a baseline, little research exists that comprehensively compares many classification techniques across multiple datasets. This thesis finds that intelligent techniques such as Neural Networks, Genetic Programming and Support Vector Machines outperform statistical techniques such as Discriminant Analysis and Logistic Regression. The second challenge is that the desire of researchers to compare results has resulted in the use of the same cross-section of factors, with little analysis as to whether or not the factors being used are impacting on the classification accuracy of the method. This thesis finds that an objective factor selection methodology leads to performance gains. The third is that far less research exists that considers whether share market or macroeconomic data can have a positive impact on classification accuracy. While this research did find some performance gains when including share market information, the difficulty of linking financial information with share market information leads to data loss that outweighs the small performance improvement. The fourth is that while most classificatory research on this problem focuses on the accuracy of the technique, less attention is given to whether the subjective clustering methods used (e.g. by “industry”) are effective, and this research finds that an objective clustering technique improves classification accuracy. Furthermore, this research builds on the existing cluster visualisation methods by developing a new and more effective cluster visualisation algorithm. Finally this research attempts to contribute to the theoretical understanding of corporate failure by analysing the classificatory surface of the resulting predictive models and performing a case study analysis of failed companies. In doing so, the model’s strengths and limitations are discussed and some of the causes of failure from the literature are identified. In summary, this research makes the following contributions to the field of bankruptcy prediction: a literature review of notable bankruptcy prediction research, a comparison of popular classification techniques, the development and testing of a new objective factor selection methodology, an examination of the effect of share market and macroeconomic data on classification accuracy, the development and testing of a new cluster visualisation method that overcomes limitations in existing methods, an examination of the effect of objective clustering on classification accuracy utilising the new visualisation method, and a case-study analysis on selected failed companies that relates the reasons for failure outlined in secondary sources to the company’s failure prediction trajectory.