Insolvency prediction by deep learning

To improve credit risk management, there is a lot of interest in bankruptcy predictive models. Academic research has mainly used traditional statistical techniques, but interest in the capability of machine learning methods is growing [1][2][3][4]. This Italian case study pursues the goal of developing a commercial firms insolvency prediction model. In compliance with the Basel II Accords, the major objective of the model is an estimation of the probability of default which is the likelihood of a default over a given time horizon, typically one year. The dataset consists of absolute values as well as financial ratios collected from the balance sheets extracted from Aida database of Bureau Van Dijk, Moody's Group, including 139,646 Italian firms. After the pre-processing procedure, 14.966 Italian micro-small firms have been selected: 13,846 ongoing and 1,120 bankrupted. The final list has been sorted out removing all records exposing missing data or negligible values. The volume of data processed places the research on a scale similar to that used by Moody's in the development of its rating model for public and private companies, "RiskcalcTM". The study has been conducted using some advanced techniques in Machine Learning and Deep Learning, such as Random Forest, Gradient Boosting, Recursive Neural Networks (RNN) and Convolutional Neural Networks (CNN). We compared several models with respect to their performances on the basis of the misclassification error (max accuracy) and the sensitivity (number of captured bankruptcy). The results have been compared also with those obtained in other studies, showing the interest of our proposal.