The default of leasing contracts prediction using machine learning

Abstract

Automatic decision support systems focused on credit risk assessment based on scoring or similar methods are often used by financial institutions. Default prediction is a very important issue. This problem is often analyzed by researchers, but products offered by banks are mainly considered. However, additional attributes of fixed assets should be used for developing the method for default prediction of leasing contracts. The aim of this paper was to develop and compare machine learning methods to increase the level of default prediction in leasing companies. We focused mainly on Random Forest Classifier, AdaBoostClassifier, GradientBoostingClassifier, and Deep Neural Networks. The main contribution is the comparison of different developed machine-learning methods using data acquired from leasing companies. The results of the experiments show the precision of the prediction of defaults was from 72,8% to 75%, and the recall of the prediction of defaults was from 76,3% to 81,6%. The precision of non-default prediction, in turn, was from 71,9% to 79,1%, and the recall of prediction of defaults from 69,7% to 71,9%. The conclusion is that the models proposed in this research may be a helpful tool in the current operations of various financial institutions. Although the functionality of the models was presented in the example of a selected leasing company, due to their flexible and universal nature, they can be successfully used in the activities of other financial institutions.