The bias of unhealthy SMEs in bankruptcy prediction models

Purpose – This paper aims to construct a financial health indicator to define the degree of financial health in order to decontaminate the estimation sample and to make predictions that are not biased by unhealthy firms. Design/methodology/approach – The binomial logit model is used to examine the likelihood that a firm will go bankrupt. In order to evaluate the accuracy of the estimated models, measures proposed by the Basel Committee on Banking Supervision are applied: cumulative accuracy profile (CAP) and the receiver operating characteristics (ROC). Findings – The proposed financial health indicator permits the heterogeneity of the firms to be reduced as well as identifying a strong firm sample to estimate the bankruptcy probability accurately. Originality/value – A drawback of all bankruptcy prediction models comes from the fact that bankruptcy is an example of a homogeneous observable qualitative response while non‐bankruptcy would be expected to be represented by a healthy firm. However, the non‐bankruptcy firms are heterogeneous and their actual probabilities of bankruptcy are non‐observable. The article adds to the previous literature on SMEs' bankruptcy prediction by using a financial health indicator to construct the estimation sample and to make accurate bankruptcy predictions.