Credit Scoring Modelling: Machine Learning Based Recommendation System For A Microfinance Organization

Over the years, massive empirical research has been carried out in the field of credit score systems construction and application for various types of credit organizations and in different economies. Recently machine learning methods including ensemble training techniques have proved an effective tool for classification problem for the purposes of credit scoring. Though the practice of credit policy making requires from the credit organization to choose a single evaluation technique. A chosen method should be able to be applied and relied on for a rather long period of time and corresponds both to the characteristics of the potential clients and the strategic goals of the organization. The present research aims to create a tool for construction of credit policy for a microfinancial organization which works with individual consumers and short-term loans. It also should take into consideration unreliable clients management costs. In this paper we suggest using multiple machine learning classifiers (including various ensemble training methods) to construct a map of alternate credit score strategies. It can be applied to compare the various techniques and choose the most relevant ones to the strategic goals of the credit policy making for the given organization. Based on the data available from the credit history of more than 13 thousand clients of a microfinance credit organization in Russia. We prove that random forests and neural nets outperform support vector machines and generalized linear models in predicting the clients reliability. We also prove that the stacking delivers significantly better prediction quality and suggest more effective set of credit policy strategies for the organization. However, this also suggests that the final recommendation about the credit policy choice highly depends on the criteria of the credit organization performance which needs additional estimation depending on the organization goals and current performance.