The evaluation of bankruptcy prediction models based on socio-economic costs

Corporate bankruptcies often have severe consequences on all stakeholders, from financial stakeholders losing their investment to employees losing their jobs. Yet traditional bankruptcy prediction models typically focus solely on predicting the event of bankruptcy itself, and do not consider the socio-economic consequences of their prediction. Therefore, this study aims to integrate these perspectives into the machine-learning (ML) modeling process to consider different costs caused by bankruptcy. We improve upon existing bankruptcy prediction models by actively taking the social and financial impacts of bankruptcy into account. Specifically, we consider two alternative evaluation metrics: the financial costs of bankruptcy, and the social impact measured using number of lost jobs as proxy. We compare a variety of machine-learning models as well as multivariate discriminant analysis and logistic regression, the latter serving as a benchmark to show the improvements that can be achieved using ML models. We apply the models on a large real-world data set from the Compustat database, containing listed companies in North America for the period from 1985 to 2020, with over 190,000 company-year observations. Our results show that small differences in statistical performance can translate into large differences regarding socio-economic costs, and that the selection of the 'best' performing model crucially depends on the evaluation metric considered.