Assessing How Bankruptcy Prediction Models Are Evaluated

Abstract

Since bankruptcy prediction became a popular research topic in the mid-1960s the model used for evaluating the research has remained largely unchanged.  A matched-pair sample design combined with a dichotomous classification test has been the standard.  This is quite useful from an academic perspective, but totally useless for any practical application.  When evaluating the paired-sample, the assumption is that random chance will correctly classify 50% of the companies as bankrupt or not, and any model that exceeds this is doing better than chance.  In the real world, of the 10,000 companies that trade on exchanges, only 600 will go bankrupt - this is a 6% failure rate, so any model should do better than 94% accuracy, not 50%.  Some companies might be eligible for bankruptcy, but choose not to file.  They might instead threaten to file, and negotiate concessions from creditors.  If this company was classified as bankrupt, would that be a correct or incorrect prediction?  The classic evaluation model would classify that as a miss, but is it really?  This paper addresses the shortcomings of bankruptcy prediction evaluation models and suggests that bankruptcy is better represented as a continuum, rather than a dichotomous situation.