Applied Research On The Prediction Of Credit Card Default By Statistical Learning Theory

In the last ten years,China's credit card market has been developing rapidly.The total number of cards issued,the total amount of credit extended and the number of card per capita by financial institutions have reached new highs.The credit card business has become one of the most important income components for banks and other financial institutions.In order to seize the market as soon as possible,a larger number of card issuance and a wider business coverage have become the goals pursued by all banks.Therefore,Banks regard the card issuance as a maj or Key Performance Indicator(KPI),and make the examination of applicants easier to get approved.Therefore,since 2010,total outstanding credit card loans to domestic Banks rose nearly tenfold,and the problem of credit card default is particularly serious.However,currently,the method to manage credit card risk hold by financial institutions is imperfect coverage and is lack of early warnings and solutions of the potential default risk users.As a breakthrough point,this paper uses the datasets provided by UnionPay which consist of some credit card transaction data from an Internet financial enterprise,to forecast the occurrence of credit card defaults.This article gets rid of the dependence on historical credit data in the past research and use the feature engineering techniques to fully tap the key information of the real transaction data accumulated by banks,we construct several forecasting models based on the statistical learning theory and prove the feasibility and availability.Finally,we investigate the connection between credit card defaults and the features.Specifically,it includes the following aspects:Firstly,this article introduces the main features contained by the credit card transaction datasets and gets eight types of transaction features by correlation calculation and feature engineering.At the same time,we divide the transaction features into the short-term,mid-term and long-term parts in order to find the connection between defaults and time factor.Also,this article introduces the main evaluation indicators.Secondly,we use logistic regression,support vector machine,random forests and gradient boosting tree to construct forecasting models.Then we make a comparison between the models based on the origin datasets and the balanced one built by undersampling techniques.Finally,we investigate the connection between defaults and transaction features by calculating the feature importance of RF and GBDT models.We also take into account the effect of noise to the forecasting performance and rebuild the models with the most important features so as to eliminate noise caused by high dimensional features.Through the above empirical results,this paper finally draws the following conclusion.Based on the credit card transaction features,we can build a default forecasting model which is feasible by statistical learning theory.But the imbalance of transaction features has a negative influence on the forecasting performance,we may construct a balanced dataset by undersampling techniques and rebuild the models in order to improve the forecasting performance.Comparing all the statistical learning theories,GBDT is the most outstanding model and is robust in the unbalanced datasets.While researching on the connection between defaults and transaction features,we find that encashment,online shopping,installment and irregular transaction time are the main cause of credit card defaults.Also,we notice that short-term and mid-term features play a more significant role than long-term features in default analysis.Surprisingly,forecasting models based on the four types of features can keep most of the performance and show better robustness.