Comparing statistical methods and artificial neural networks in bankruptcy prediction

• Main Author: Chu, Jung
• Published: University of Warwick 1997
• Subjects: 658; HG Finance : QA Mathematics
• Online Access: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.232828

The use of multivariate discriminant analysis (MDA) and logistic regression procedure (Logit) in predicting business failure has been explored in numerous studies since 1960s. Recently, a newly developed technique, artificial neural networks (ANNs), has attracted much attention and has been applied to bankruptcy prediction area. At the same time, many papers attempted to compare the predictive ability of these two distinct classes of discriminators in order to find a best failure prediction method. However, most of their results, despite showing the superiority of ANNs, have been sharply criticised either for the unfair comparison or for their specific data selection. There is a need to undertake theory-based research to identify problem characteristics that predict when ANNs will forecast better than statistical models; to identify which input variable characteristics predict when ANNs will improve model estimation; and to identify when this advantage would give substantially improved forecasting performance. Motivated by the limited amount of research on investigating the relative effectiveness of traditional methods as compared to the ANNs under a wide variety of modelling assumptions, one of the objectives of this study is to compare their classification capacities on a theoretical basis, and to evaluate the robustness on certain situations through the simulation study. The investigation is conducted on two popular statistical techniques—the MDA and the Logit, as well as two different learning algorithms of ANNs—the standard generalised delta rule (GDR) and the Projection approach (Proj). This can be regarded as the horizontal assessments of bankruptcy prediction. The other aim of this thesis is to evaluate the impacts of variations in failure prediction models through the empirical study. These variations involve the issues we often encounter in the real world, such as the different sizes of sample, a choice-based sampling bias, the sensitivity of optimal cutoff points to misclassification costs of Type I and Type II errors, and the imbalance of the composition of failed to nonfailed firms between training and testing data sets. This can be viewed as the vertical assessments of bankruptcy prediction. The simulation results indicate that the neural networks are indeed competitive approaches on bankruptcy prediction. In particular, the Projection network, which was developed to overcome the drawbacks that a commonly used GDR backpropagation algorithm often experiences, proves its remarkable superiority not only quantitatively (i.e., lower overall accuracy), but also qualitatively (lower Type I and Type II errors). The Projection network holds a promise for future elaboration. Moreover, the outcomes of empirical experiments enhance our knowledge of some factors in constructing a failure forecasting model. This knowledge is related to both traditional statistical tools and modem neural networks and is essential for decision making.