Threshold Moving Approach with Logit Models for Bankruptcy Prediction

This article focuses on the issue of the classification capability of logistic regression models in the area of bankruptcy prediction within two manufacturing sectors. Most authors undervalue the setting of a threshold for classification and use a standard dividing point. However, the results of this article show that for data that truly reflect the market situation, this standard threshold is inappropriate, as it leads to a high classification error for bankrupt companies, which are less represented in the dataset than active (healthy) companies. In order to find a suitable threshold, two criteria derived from empirically estimated ROC curves were used in this article, which made it possible to balance the error rate within the group of active and bankrupt companies.

Automated summary

This article examines the classification capability of logistic regression models in bankruptcy prediction within two manufacturing sectors. It suggests that using a standard threshold leads to high classification error for bankrupt companies and proposes using empirically estimated ROC curves to find a suitable threshold.