Machine learning approach as an alternative tool to build a bankruptcy prediction model in banking industry

Abstract

This paper focuses on the study of a bankruptcy prediction model using a hybrid machine learning that combines two synergistic algorithms i.e. two-class boosted decision tree and multi-class decision forest. The hybrid model ensures the building of multiple decision trees whereby the latest tree corrects the previous tree, learning from the tagged data and subsequently votes on the most popular tree as the final decision of the ensemble. This hybrid machine learning is proposed to be an alternative of the bankruptcy prediction models that is able to produce three major classifications i.e. bankruptcy, grey area, and non-bankruptcy. There are five variables considered in the hybrid model which consist of working capital for total assets, retained by total asset, earnings before interest and taxes on total asset, market value of equity to total bank value of liabilities and sales of total asset. These input data are applied and tested to the public dataset produced by Bank Indonesia from year 2011-2015. The hybrid model shows a significant result whereby the overall area under curve (AUC) had successfully achieved 95% value that indicates the capability of the hybrid model to train the test data and identify the relationship of input-output data. This finding suggests that the machine learning approach can be treated as an alternative tool to build a bankruptcy prediction model for banking industry. Introduction