Predicting and Interpreting Business Failures with Supervised Information Geometric Algorithms

Business failure belongs to the most investigated topics in the business literature (Huang et al. (2008)). A plethora of works have addressed the problem using conventional statistical or machine learning techniques that are known to suffer from distributional assumptions, representational bias, statistical inconsistencies or over-fitting in generalization, and therefore bring results that have to be read with most caution for risk-free use (Crutzen and Van Caillie (2008); Ravi Kumar and Ravi (2007)). In this paper, we analyze corporate tax return from a thousand French companies using a powerful blend of wavelet-based data modeling and new classification techniques born from information geometry that do not suffer any of the former drawbacks (Nock and Nielsen (2009)). It is quite remarkable that particular cases of these techniques have been known for decades, yet researchers in the business failure field claim that they have remained under-exploited (Ravi Kumar and Ravi (2007)). Our results are a clear-cut advocacy for the usefulness of these techniques in this field, as they display the ability (i) to produce models accessible even to green users, whose interpretation sheds light on the statics and dynamics of business failure, (ii) to outperform in generalization classification algorithms traditionally used in the business failure field by orders of magnitude. JEL Classification Numbers: C38 — Classification Methods; C53 — Forecasting Models.