关键词: 贷后风险管理, 个人信用风险, 消费贷, 遗传算法, BP神经网络

Abstract: In recent ten years, with the change of consumption habits of young groups, the concept of advanced consumption has been deeply rooted in people’s minds, and the consumer credit market has also achieved considerable development. By the end of 2020, the outstanding balance of bank card loans had grown from less than 500 billion yuan in 2010 to 7.91 trillion yuan, and the outstanding balance of consumer loans had risen to 50 trillion yuan. Young low-income people between the ages of 18 and 35 are rapidly becoming the new blue sea of consumer finance, and the problems of overdue and default risk become increasingly prominent. The post-loan supervision of consumer loans is becoming a new issue that needs to be solved urgently. Compared with mortgage loan and car loan, consumer loan has significant characteristics of small amount, high frequency and no collateral. The existing risk management research focuses on personal credit identification before credit granting, while the literature on post-loan risk management is still relatively rare. With the massive development of consumer loans, personal credit risk management after credit granting is becoming a new problem faced by banks. The traditional model that only relies on default rate, overdue times and other post-index is difficult to achieve effective control of post-loan risk.
From a data-driven perspective, this paper combines genetic algorithm and BP neural network to establish a consumer loan post-loan risk grade evaluation model, then compares the performance differences of customer groups with different risk levels. The data comes from the consumer loan, 230,923 flow record of a bank in western China in 2019, involving 199,603 consumer loan customers. The results show that the post-loan risk grade evaluation model based on GA-BP neural network can effectively classify customer credit risk grade. High-risk V customers that should be monitored mainly account for 0.037%. Higher risk customers (Grade IV) account for 0.61% of customers. Risk I, II and III customers with very low default rates account for more than 99 percent. In other words, in the process of post-loan risk monitoring, only 0.637% of accounts can be monitored in real time to effectively reduce the overall delinquency rate and default rate, thus significantly reducing the cost of post-loan risk monitoring. Further comparison of the performance differences of customer groups with different risk levels shows that although the subjects have passed the pre-loan personal credit screening, their post-loan risks show significant differences. The behavioral differences of customer groups with different credit risk levels can be reflected from three dimensions: From the fund flow dimension, the more active the deposit of funds, the stronger the account liquidity, because of the lower credit risk. The higher the spending level, the higher the credit risk. From the income level dimension, the income level of customers with different credit risk levels presents U-shaped distribution. The more stable the income is, the less credit risk the customer has. From the repayment tendency dimension, the higher the minimum repayment ratio, the lower the credit risk has, while the higher the expenditure ratio, the higher the credit risk is. The management revelation is that hierarchical control can effectively realize the dynamic monitoring of the borrower’s consumption behavior after credit granting and timely prevent the post-loan credit risk.

Key words: post-loan risk management, personal credit risk, consumer loans, genetic algorithm, BP neural network