Internet loan risk control system includes

Internet finance is only ten years long. Although the time is not long, it cannot be said that the scale of its influence has undergone earth-shaking changes. This can be seen from the name. Internet finance is widely known, not the financial internet that many experts and scholars once discussed. Traditional financial internetization is not enough to describe the great changes that have shaken China's financial system. Internet finance, of course, is not limited to credit, but also includes new formats such as banking, securities and basic information. It is a veritable "showstopper of the financial industry".

but in any case, the initial debate between internet finance and financial internet is around credit. We focus on the topic of credit. How does the credit risk control system adapt to the changes of the times? Today, we discuss the common model strategy systems for risk control and their second-order problems: Why is it such a system?

The Bao Yue, which was born in p>213, made the old tree of the Monetary Fund blossom and set off the whole market. Otherwise, Internet finance will only be a quiet sub-topic in the IT circle. This is why the industry generally defines 213 as the first year of Internet finance.

In p>214, the first Internet consumer financial product "Baitiao" was established, and the first Internet bank "Weizhong Bank" was approved by the regulatory authorities.

at this time, the mutual loan business is still in its infancy. What can risk control do? Under the new model, there are no users, the performance cycle of traditional bank credit is one year or even several years, and there are even fewer standard data products, not to mention the big data model, and the statistical model can do nothing. At this time, the credit can only go to high-quality people, such as credit card customers, such as some advanced categories of e-commerce trading customers, and so on.

at this time, the risk control can be said to be based on customer risk control or white list risk control. This leads to the crowd-based Internet consumer finance, including college students' staging, blue-collar/white-collar staging, farmers' staging, and then the scene-based Internet consumer finance, such as rental staging, home improvement staging, tourism staging, education staging, medical beauty staging and so on.

Then, with the rapid development of P2P, online lending and other mutual financial services, the main body of consumer finance market is further enriched. At the same time, the all-round development of the mobile Internet has led to a surge in data and information. To some extent, data has become a strategic resource as important as oil. Under the rising trend of big data, a large number of data companies have been established. The development of computing and big data has triggered a frenzy of machine learning, and the level of risk control has been greatly improved.

In the barbaric period of 17 or 18 years, many people dared to borrow money as long as they had a little money and courage. Access to several tripartite data sources, set some general rules, make a credit model, and find a few people to collect it, and you can start the flow.

at this time, the complete technology of risk control is quite mature. The tripartite data covers multiple dimensions, such as credit reporting, UnionPay, operators, public security, justice, industry and commerce, taxation, bulls and so on. What samples to take, what labels to set, what data to access, what algorithm to use, and what pass rate to make a model strategy that matches the product are not difficult. What really matters is the product itself, that is, the amount, pricing, number of installments, repayment method, etc.

Later, a series of regulatory policies were introduced one after another, and the industry moved from high-speed development to a new path of standardization and rectification. After the storm, how to remain stable and far-reaching in a healthy competitive environment and development space has become a new proposition, and the industry has entered and will continue to be in the era of refined operation.

at this time, it is a new proposition that how to make the risk control detailed and how to serve the needs of business forms.

The changes in the Internet era characterized by product changes seem to be more striking.

Time product manager's job may be to draw prototypes, so it is very powerful to make the interaction to the extreme; Then there will be strategy optimization and functional evolution. The product managers of ToC business are all strategic product managers in terms of work content, and they also need to know the data. Now and in the future, it has quietly become a plan combing under the business form and organizational relationship.

speaking of which, I finally have a background. The industry is changing and the risk control system is also changing.

H

In the past few years, there have been so many new practitioners in the credit industry, or showstopper, so many new online credit products and the risk control games behind them.

but to sum up, there are only three common reliable model strategy systems.

the first one is based on rules and supplemented by models. Most of them appear in the early stage of the exhibition industry, with few samples and data. Risk control mainly depends on the experience of experts, as do cash loans without scenes and consumer loans with scenes.

in the early stage of the exhibition industry, or the cold start stage, samples need to be accumulated gradually, and risks need to be exposed for a long time, so credit must be strictly enforced. Before we know the risk results, the standard of strict admission or loose admission is the pass rate, so the initial pass rate is generally low.

what is the pass rate of the card? Maybe you have a scene. Of course, a scene data card is ok, but it is not enough. The risk is lagging behind, after all, we have to access some tripartite data. According to these data, according to your scene, according to your product positioning, formulate corresponding policies and rules.

Then, with the repayment date coming, a group of users begin to implement it, and the data of new households will definitely lose money on the score, so you will want to optimize it. You know that risk is closely related to your product and your supporting operation. You will feel that the standard product is not enough, and you have to model it yourself.

under the condition of limited samples, too much data access is meaningless. At this time, the model can only be a small data model rather than a big data model. If there are no major changes in your products and there are no mistakes in model development, this model may bring some gains. With the increase of business scale, the model is frequently updated and optimized.

but at this time, the whole risk control system is dominated by rules, supplemented by models. White list, blacklist, age and geographical restrictions, bulls, public security and judicial information, standard product scores. Your risk control is based on the general expert rules in the industry, and the model is only an aid. To put it bluntly, your sense of security comes from various rules.

this system not only appeared in the early stage of the exhibition industry, but also many platforms have been in this stage for a long time, or even died. Its representative scenario is payday loans, and a series of online loans derived from it, which have high risks and make up for capital losses through high pricing and short cycle.

the second one is model-oriented, supplemented by strategy. After the samples are relatively rich, the importance of the model will gradually increase, especially after the intervention of the industry and the sinking of customer quality, accurate identification becomes necessary.

In addition to the empirical series of rules among the initial experts, some variable rules are constantly adjusted. After these rules, how can we make more refined decisions according to the model?

one way is to make the model as effective as possible, and then everyone can further filter through this model. In order to be effective enough, the complexity of the model will become high, and the defect is that you can't tell why this person was passed and that person was rejected. No matter how the explanatory tool works after doing something, this defect can't be made up in essence.

Another method is to make the model effective enough and interpretable enough. In order to pursue interpretability, we can classify data, that is, features, and build corresponding models based on each type of features, such as overdue models, such as multi-head models, such as transaction models, and so on. The feature combination of the same dimension retains a certain degree of interpretability of this dimension. Through each model in series, it can be judged that the user is not good at overdue performance, or the bulls are serious, or the UnionPay transaction score is risky.

in this kind of risk control system, your sense of security comes from the model. As long as the AUC and KS of the model remain high, you will feel more at ease no matter how much money you lend every day.

in this kind of system, the model is very important, and the strategy is differentiated by the model. A model-based risk control team will eventually move towards this approach, and many banks will also move towards this approach. Because real strategic talents are scarce.

the third one is based on strategy and supplemented by model. Pay attention to the difference between the first one and the first one. The first one says that rules are the mainstay, and here we say that strategies are the mainstay. Rule-based rules are simple, universal, empirical and serial.

at this stage, the samples are very rich, and the user data is continuously mined in the scene, and all kinds of effective tripartite data are accessed. The big data model has a good effect and is constantly pursuing better. The model is very important, just as the tool of "if a worker wants to do a good job, he must sharpen his tool first" is very important.

at this time, the model acts as a tool for policy. The model can not pursue interpretation, and the strategy is the main one, and the strategy keeps the decision interpretable.

The decision should be interpretable, because the future may not be as it is now, and we can't bear the harm of extreme situations. Just like investing, you can digitize a set of decisions to avoid all bear markets and find all bull markets, but there are only a few examples. Do you dare to use them now? Decision-making must be as simple as possible. It can make mistakes, but the mistakes should be small and the benefits should be great.

the essence of the strategy is clustering, age is clustering, income is clustering, bulls are clustering, models are clustering, and risk is clustering. No matter whether it is credit granting or loan management, no matter what you do to users or what you want users to do, you should distinguish users.

the sign of this kind of system is that there are many important subgroups in the decision-making system, that is, decision branches, and the model is used as the final guarantee. As a tool of strategy, the model can be used as much as it needs, and a tool can be used in a large range or a small range.

The decision branch means that the use of the model by the policy is not uniform, and not all users with less than 6 points will be rejected.

Your sense of security comes from strategy, and more specifically, it is strategy grouping.

under the strong supervision of the industry, non-licensed institutions are constantly falling down, and only giants can barely survive. More and more risk control is this system. Nothing else, just refinement requirements.

our discussion above has covered the development background of these systems, and the risk control systems are different in different stages, platforms and business scenarios.

why has such a system been developed? It seems that the answer to this second-order question is data. The number of samples, the number of data dimensions and the number of features determine the relationship between the underlying model and strategy.

but this is not the essence of the problem. I think the essence of the problem lies in the risk-reward ratio.

If you play the game like "714" anti-aircraft gun, it is the most profitable to set up a few rules and let the money go out for interest. The cycle is short and the liquidity is strong. The annualized interest rate is several times or even ten times, and the risk is a principal, which is nothing. There are no high-quality people in the target customers, so what else can we do to accurately identify the model?

when this game is no longer legal, the risk return is getting lower and lower. Generally, the bank's customer pricing is within 18% per annum, and the consumer finance company is basically in the early 2%. If you want to make profits within this range, you must keep the risks stable and controllable, and the service target should be targeted at high-quality customers. Without a big data model, it is impossible. Of course, there is no shortage of high-quality traffic giant optical white list is enough.

later, the traffic became more and more expensive, and customers were constantly infiltrated by multiple platforms, so it is very important to make a good stock. Save a flow to calculate a flow, increase a little balance to calculate a little balance, and start to dig every user to the extreme. Customer grouping and refined operation have become the current trend.

just, is this kind of mining and operation what customers need?

if you look up to where you are going, it's always bumpy. Looking back, all roads are smooth.

The title map is from Unsplash, based on the related Q&A of CC protocol: