Alternative methodologies in studies on business failure: do they produce better results than the classic statistical methods?

Over the last 35 years, the topic of company failure prediction has developed to a major research domain in corporate finance. Academic researchers from all over the world have been developing a gigantic number of corporate failure prediction models, based on various types of modelling techniques. Besides the classic cross-sectional statistical methods, which have produced numerous failure prediction models, researchers have also been using several alternative methods for analysing and predicting business failure. To date, a clear overview and discussion of the application of alternative methods in corporate failure prediction is still lacking. Moreover, frequently, different designations or names are used for one method. Therefore, this study aims to provide a clear overview of the alternative research methods, attributing each of them a fixed designation. More in particular, this paper extensively elaborates on the most popular methods of survival analysis, machine learning decision trees and neural networks. Furthermore, it discusses several other alternative methods, which can be considered to have a certain value added in the empirical literature on business failure: the fuzzy rules-based classification model, the multi-logit model, the CUSUM model, dynamic event history analysis, the catastrophe theory and chaos theory model, multidimensional scaling, linear goal programming, the multi-criteria decision aid approach, rough set analysis, expert systems and self-organizing maps. This paper discusses the main features of these methods and their specific assumptions, advantages and disadvantages and it gives an overview of a number of academically developed corporate failure prediction models. Several issues viewed in isolation by earlier studies are here considered together, which is of major importance for gaining a clear insight into the possible alternative methods of corporate failure modelling and their corresponding features. A second aim of this paper is to find an answer to the question whether the more sophisticated, alternative modelling methods produce better performing failure prediction models than the rather simple classic statistical methods. The analysis of the conclusions of a large number of empirical studies comparing the classification results and/or the prediction abilities of failure prediction models based on different techniques seems to indicate that we may question the benefits to be gained from using the more sophisticated alternative methods.