Do you understand the cash loan risk control model?

Recently, the supervision of cash loans has been imminent. On the one hand, the actual borrowing interest rates of many small and medium-sized cash loan platforms are too high; on the other hand, the risk control of some platforms' cash loan business is more like "the emperor's new clothes", so that the bad debt rate of the entire industry remains high.

From a regulatory perspective, in order to continue operating, the cash loan business will have to significantly reduce loan interest rates and reduce various handling fees. Therefore, improving the level of risk control and reducing bad debt losses has become the only way to ensure the operating profits of the cash loan business.

1. A brief analysis of the cash loan risk control system: point-surface-point process

The first "point" refers to the starting point. The design of the cash loan risk control system needs to take the product itself as the starting point. Cash loan products have nothing more than four elements: interest rate (including various rates), term, amount, and target group. For each type of target group, they have certain patterns and characteristics in terms of liquidity needs, expected future cash flow, consumption concepts, income levels, and credit status, which in turn affects their application quotas and loans. Interest acceptance level, repayment ability and willingness to repay, etc. Therefore, rationally designing products can effectively reduce the difficulty of risk control while maximizing revenue. For example, for white-collar workers, the monthly salary feature is more suitable for loan periods of one month or less.

In addition, in addition to the characteristics of the cash loan product itself, its promotion channels are also quite important. If a large number of non-target people are introduced through a certain promotion channel, this will not only reduce the efficiency of promotion costs and the judgment accuracy of the later risk control process, but also generate a large amount of biased data, which is not conducive to the iterative upgrade and improvement of the risk control model. Product redesign.

“Aspect” refers to the specific risk control process. Distinguished by time period: the risk control process includes three stages: pre-loan, loan, and post-loan. The pre-loan stage is the core stage of the entire risk control process. This stage includes three steps: application, review and credit granting. Figuratively speaking, the pre-loan stage is a stage of filtering out impurities. The third-party credit data, blacklists, anti-fraud rules, and risk control models are layers of filters with different apertures. The loan stage is mainly about tracking and monitoring the borrower's personal information. Once abnormal information is generated, risk control personnel can promptly discover and contact the borrower to ensure the safety of the loan as much as possible. The work in the post-loan stage focuses on collection. In addition, if the borrower applies for extension or loan renewal, it needs to be re-examined at this stage based on historical data, behavioral scorecards, etc., and corresponding limit adjustments and risk pool management must be made. Throughout the entire risk control process, the concentration of borrowing needs to be properly managed to prevent insufficient liquidity caused by concentrated borrowing and concentrated overdue loans.

The second "point" refers to the key point. The entire cash loan risk control system focuses on two aspects.

First, anti-fraud. Compared with the risk control under the traditional lending model, cash loan risk control is a light risk control. Due to its small-term, short-term characteristics, cash loan risk control pays more attention to the borrower’s willingness to repay rather than its ability to repay. Moderate overdue fees will not only not affect the normal operations of the platform, but can actually increase its revenue through overdue fees.

Therefore, anti-fraud is the primary topic of cash loan risk control. At present, online loan fraud includes intermediary agency, gang crime, machine behavior, account theft, identity fraud and series transactions, etc. For these fraudulent behaviors, commonly used anti-fraud rules include comparison, cross-checking, strong feature screening, risk relationships, and user behavior data analysis.

Second, identification of long-term lending behavior. Multi-lending refers to the same borrower having loans from multiple lending institutions. Currently, the identification of long lending behavior includes two aspects: (1) Obtaining long lending data. Since most of the target groups of cash loans are long-tail groups that are not covered by traditional lending institutions and lack complete central bank credit data, some platforms engaged in cash loan business will cooperate with each other to achieve full sharing of loan application data. . In addition, the cash loan platform will inevitably leave a large amount of identity information of the loan applicant when the third-party credit agency inquires about each loan application record.

After this part of information is filtered by the query anomaly detection algorithm, a reliable long lending database will be formed. (2) Identification of vicious long lending behavior. Vicious long lending behavior refers to lenders borrowing new funds to repay old loans or having large amounts of long loans during the same period. The identification of new borrowing and repaying behavior can be combined with the loan application interval and loan period. If the loan application interval is significantly shorter than the loan period, it means that the loan application has a greater risk of borrowing new money and repaying old money.

2. Challenges: Contradictions and changes

1. Diversification, technology, and Internet of fraud methods

Fraud and anti-fraud have always been about lending One of the main contradictions in the industry. With the rapid development of online loan business, online loan fraud based on information technology has also become more and more intense. When scammers also start to play with big data and machine learning, it is conceivable that many risk control personnel are heartbroken.

For example, mobile phone verification is currently one of the most commonly used online review methods. It includes two forms: SMS verification code and filling in the operator service password. But this method is also an opportunity for fraud gangs. Because they have a kind of technical equipment - cat pool. To understand it simply, it is a simple mobile phone with the function of sending and receiving text messages and having "n cards and n standby". One computer can connect to multiple MaoPools, and one MaoPool can insert 8-64 SIM cards. Along with this, there is the so-called "card collection" and "card maintenance" business. When the number time reaches a certain standard, it is possible to use mobile phone verification as an anti-fraud method.

In addition, the use of some emulators can help fraudsters easily modify the IMEI, MAC, IP, GPS and other device and environmental information of the mobile phone. Under these layers of camouflage and packaging, anti-fraud methods that use device and environmental information seem a bit feeble. Moreover, some personal information, such as ID card information, social media accounts, bank card accounts and even USB shields, can be purchased by fraud gangs online or searched using search engines. Many times, some anti-fraud methods are effective not because they cannot be cracked, but because the cost of cracking is high, causing fraud groups to give up this method.

2. Cold start of the risk control model

“Cold start” is the primary problem faced when building a big data risk control model. Especially for some start-up cash loan platforms, data accumulation is a process starting from scratch. In the early stages of accumulating data, huge costs are bound to be incurred. On the one hand, in addition to ensuring normal risk control processes, platforms also need to invest a lot of labor costs in collecting data, building models, and data backtesting; on the other hand, platforms have to invest high capital costs in purchasing third-party data. Compared with the total number of long-tail user groups of nearly 1 billion that are not covered by the central bank's credit data, the current customer base for cash loans is still limited, and most platforms are facing the "cold start" problem.

The currently commonly used method to solve the data cold start problem is to start with external data. Due to the lack of historical credit records and personal credit data of loan applicants, the risk control model loses the basis for directly considering the borrower's default risk. Therefore, if external data such as user behavior can be combined with collaborative filtering algorithms such as Eigentaste, people with a higher risk of fraud can be identified to the greatest extent and filtered. However, the current status quo is that most platforms lack the power and ability to process external data, and often use manual review supplemented by some simple anti-fraud rules.

3. The contradiction between user experience and anti-fraud

In the eyes of cash loan users, user experience is reflected in the speed of borrowing and the ease of application. However, anti-fraud requires users to provide a variety of personal information, which greatly reduces the quality of user experience. In the past, many platforms pursued the principle of “high returns covering high risks” and overemphasized traffic. Moreover, there are various cash loan products on the market. In order to ensure traffic, many platforms have posted slogans such as "only ID card and mobile phone number are required" and "loan will be released within XX minutes after application". However, as supervision becomes stricter, the "high-yield, high-bad-debt" operating model will gradually be eliminated. In order to control bad debts, cash loan platforms have to face the conflict between user experience and anti-fraud again.

On the one hand, the platform needs to optimize the anti-fraud model, reduce the dimensionality of the entrance data as much as possible, and shorten the review time of the risk control model; on the other hand, optimizing the user experience from other perspectives such as customer service and ease of repayment is also a way to alleviate user experience and feedback. One of the possible ways to deceive contradictions.

3. Development Direction

1. Use of unstructured data

The sparsity problem of structured data such as personal credit data will be a big problem in the future. It has existed in the cash loan industry for a long time. Corresponding to this is the proliferation of a large amount of unstructured data. Due to the serious leakage, theft, and trafficking of basic personal information, the anti-fraud efficiency of conventional structured data has been greatly reduced. Compared with structured data, people's behavioral data is more difficult to simulate and can more comprehensively describe loan applicants, which has a significant effect on reducing the error rate of anti-fraud models.

From the perspective of the application of unstructured data, it is difficult to unify the logic between them. The problems of data anomalies, redundancy, and missing are serious and difficult to process. Therefore, seeking cooperation from third parties such as big data companies and traditional Internet industry giants will be the first choice for small and medium-sized cash loan platforms. There are already some products on the market that serve finance by refining unstructured data, such as a commercial SMS semantic analysis service. In addition, Tencent, one of the BATs, also cooperated with Qian Niuniu to launch a purely model-based cloud risk control system - "Yuanfang". The biggest feature of this system is the introduction of Tencent’s massive social data.

2. Differentiated pricing

Differential pricing can also be understood as precise pricing. The essence of differentiated pricing for cash loans is to accurately price the credit and fraud risks of each loan applicant. At present, the pricing standards of various cash loan platforms are too simple, and they basically adopt the method of interest rate plus miscellaneous fees. Some platforms will make rate adjustments for users who renew their loans. There are also a few platforms that will refer to the personal information dimensions provided by the lender when applying. However, in general, the current pricing standards are not suitable for the "low rate" characteristics of the future cash loan industry. The so-called price advantage between platforms will be minimal. The demand for customized small loans under precise pricing may become the highlight of the platform.

The construction of the big data risk control model provides technical guarantee for the realization of differentiated pricing. Based on a large amount of network behavior data, user transaction data, third-party data, partner data, etc., through natural language processing, machine learning, clustering algorithms, etc., the model can create for each loan applicant including basic personal information, Multi-dimensional data portraits including behavioral characteristics, psychological characteristics, economic status, interests and hobbies, etc. With these dimensional characteristics and a large number of historical loan records, differentiated pricing strategies for different lenders, different amounts, and different maturities will become a reality.

Summary

Behind the industry reshuffle are the efforts of cash loan platforms to survive. How to ensure compliance, how to obtain low-cost funds, how to replace manpower with technology, and how to find a balance between risk control costs and bad debt rates are issues that the platform needs to think about and solve in the future. We believe that under the searchlight of industry policies, gold will always shine in the end.