Improving bankruptcy prediction using Data Envelopment Analysis scores

Most current bankruptcy prediction models are based on financial ratios, although their usage is not supported by formal theory and their interpretation is problematic. One of the prospects for improving the predictive models is the study of other firm performance measures, such as the data envelopment analysis (DEA) scores. However, this raises the problem of choosing the optimal DEA specification, since it determines the shape of the efficiency frontier and the predictive properties of the model. This paper presents a method for automatically designing DEA models whose scores are then used as features to improve the quality of bankruptcy predictors. The method has two goals. The first is to improve accuracy. The second objective assumes that if DEA scores improve the prediction, then the specification of this model can provide information about failures. At the first step, accounting measures that are potentially suitable for the DEA are selected using hierarchical clustering. The second step explores the causal relationships between the selected measures. The third step calculates pure technical efficiency, scale efficiency and mix efficiency. Experiments with two datasets show that the inclusion of these scores in the list of features improves the AUC-ROC by more than 20%, which is superior to previous works. The analysis of the DEA models provides insight into the reasons for a firm’s failure in both stable and crisis periods.