The Application of Machine Learning Algorithms in Credit Card Default Prediction

With the rapid development of the credit cards industry, there is an increasing number of delinquency rates on credit card loans, which imposes a financial risk for commercial banks. Therefore, successful resolutions of the risks are important for the healthy development of the industry in the long term. The existed methods, such as FICO model [1] (developed by Fair Isaac Company) can estimate the probabilities of credit card defaults, but it based totally on the subjective judgement of people. This means that FICO model comes with several undesirable problems, including low efficiency, low accuracy, time-consuming and high labor costs. In addition, the data of credit card defaults is always unbalanced, since few clients default in real world, which brings challenges to default model construction. In current big-data era, machine learning methods [2] are popular for its high efficiency and high accuracy. In this paper, we employed several classical machine learning algorithms, including logistic regression [3],decision tree [4] and ensemble learning [5] (adaboosting [6], random forest [7]), to build credit default prediction models. To solve the problem of unbalanced data [8], we further build corresponding weighted models so that it can improve the prediction accuracy of default class with slightly higher prediction error of non-default class. The results show that random forest models with weight is the best, which has achieved an accuracy of 82.12%. It achieves the goal of fast learning speed, high parallelism efficiency and high-volume data. Overall, machine learning algorithms has practical application value which can evaluate the delinquency precisely.