The predictiveness of bankruptcy models: methodological problems and evidence

This paper addresses three methodological problems in previous studies of bankruptcy models using MDA and logit/probit techniques: (i) the use of choice-based and equally distributed samples in model estimation and validation; (ii) the use of arbitrary cut-off probabilities; and, (Hi) the assumption of equal costs of errors in prediction tests. Corrected MDA and probit models are estimated for a sample of Australian companies from 1966 to 1986. In contrast to previous studies our results show that first, bankruptcy prediction models based upon publicly available financial statement infor Mation possess insignificant predictive power; and second, the variables chosen for such models are characterised by non-multivariate normality and heteroskedasticity.