The bankruptcy prediction research domain continues to evolve with many new different predictive models developed using various tools. Yet many of the tools are used with the wrong data conditions or for the wrong situation. Using the Web of Science, Business Source Complete and Engineering Village databases, a systematic review of 49 journal articles published between 2010 and 2015 was carried out. This review shows how eight popular and promising tools perform based on 13 key criteria within the bankruptcy prediction models research area. These tools include two statistical tools: multiple discriminant analysis and Logistic regression; and six artificial intelligence tools: artificial neural network, support vector machines, rough sets, case based reasoning, decision tree and genetic algorithm. The 13 criteria identified include accuracy, result transparency, fully deterministic output, data size capability, data dispersion, variable selection method required, variable types applicable, and more. Overall, it was found that no single tool is predominantly better than other tools in relation to the 13 identified criteria. A tabular and a diagrammatic framework are provided as guidelines for the selection of tools that best fit different situations. It is concluded that an overall better performance model can only be found by informed integration of tools to form a hybrid model. This paper contributes towards a thorough understanding of the features of the tools used to develop bankruptcy prediction models and their related shortcomings. © ۲۰۱۷ Elsevier Ltd. All rights reserved.

Introduction

The effect of high rate of business failure can be devastating to firm owner, partners, society and the country’s economy at large (Alaka et al., 2015; Edum-Fotwe, Price, & Thorpe, 1996; Hafiz et al., 2015; Xu & Zhang, 2009). The consequent extensive research into developing bankruptcy prediction models (BPM) for firms is undoubtedly justified. The performance of such models is largely dependent on, among other factors, the choice of tool selected to build it. Apart from a few studies (e.g. Altman, 1968; Ohlson, 1980), tool selection in many BPM studies is not based on capabilities of the tool; rather it is either chosen based on popularity (e.g. Abidali & Harris, 1995; Koyuncugil and Ozgulbas, 2012; Langford, Iyagba, & Komba, 1993) or based on professional background (e.g. Altman, Marco, & Varetto, 1994; Beaver, McNichols, & Rhie, 2005; Hillegeist, Keating, Cram, & Lundstedt, 2004; Lin & Mcclean, 2001; Nasir, John, Bennett, Russell, & Patel, 2000). This is because there is no evaluation material which shows and compares the relative performance of major tools in relation to the many important criteria a BPM should satisfy. Such material can provide a guideline and subsequently aid an informed and justified tool selection for BPM developers. Most prediction tools are either statistical or artificial intelligence (AI) based (Balcaen & Ooghe, 2006; Jo & Han, 1996). The most common statistical tool is the multiple discriminant analysis (MDA) which was first used by Altman (1968) to develop a BPM popularly known as Z model, based on Beaver’s (1966) recommendation in his univariate work. MDA, normally used with financial ratios (quantitative variables), subsequently became popular with accounting and finance literature (Taffler, 1982) and many subsequent studies by finance professionals simply adopted MDA without considering the assumptions that are to be satisfied for MDA’s model to be valid. This resulted in inappropriate application, causing developed models to be un-generalizable (Joy & Tollefson, 1975; Richardson & Davidson, 1984; Zavgren, 1985). Abidali and Harris (1995), for example, unscholarly employed A-score alongside Z-score (i.e. MDA) in order to involve qualitative managerial variables, alongside quantitative variables, in their analysis when logistic regression (LR) [or logit analysis] can handle both types of variables singularly.