Machine Learning-Based Bankruptcy Prediction for Italian SMEs in Manufacturing with Comparative Analysis of Pre- and Post-COVID-19 Periods.

Abstract

Bankruptcy or business failure poses significant threats to enterprises and the global economy. Various stakeholders, including businesses, investors, governments, and researchers, seek effective ways to predict and mitigate associated risks. Reliable prediction methods enable proactive measures to enhance business financial health or minimize economic losses. This research focuses on small and medium-sized enterprises (SMEs), vital components of the European economy. This research investigates the efficacy of one-class and two-class classification models in predicting SME bankruptcies, with a particular focus on Italian Manufacturing SMEs in 2018 (pre-COVID-19), 2019 (during-COVID-19), and 2020 (post-COVID-19), aiming to assess whether balance sheet data can effectively predict bankruptcy across different economic periods. Using a dataset of over 100,000 Italian SMEs and 57 financial ratios from the AIDA Database, the predictive performance of six machine learning models is assessed. The models include three one-class classification methods: One-Class Support Vector Machine, One-Class Logistic Regression, and Isolation Forest, as well as three two-class classification methods: Logistic Regression, Two-Class Support Vector Machine, and XGBoost. Results show that two-class classification models, generally outperform one-class models in terms of Geometric Mean and AUC-ROC.