Bankruptcy prediction via machine learning

Abstract

Bankruptcy prediction is the problem of detecting financial distress in firms, which could potentially lead to their bankruptcy. A good prediction method can allow the company’s stakeholders to take action in order to improve the business’ financial health or limit their economic losses. Given its relevance, this problem has been analyzed since the 1930s, and a plethora of prediction models have been proposed, starting from univariate statistical models to more complex, multivariate approaches (like the famous Altman Z-Score). However, although these models performed well in the context that they were applied, their predictive power decreased dramatically when used in different scenarios, making them unreliable. Since the 1990s and with the beginning of the “Big Data Era”, machine learning models have proved to be the superior choice for bankruptcy analysis, since they are more versatile and offer much better predictions. This study demonstrates the predictive power of machine learning models based on a dataset of more than 6000 Taiwanese firms and 95 financial ratios. Three models have been used: the Logistic Regression, The LASSO Regression and the Random Forest, which, after being tested and evaluated, proved their effectiveness in predicting bankruptcy.