Shulou(Shulou.com)06/02 Report--

Author: Zhou Zhou 2020-06-18 10:13 introduction: what financial pain points can not be solved by AI risk control?

There is no truth that can solve all the problems in life, and no technology that can help an enterprise to be invincible.

In the financial technology industry, even if it is a simple project, it still needs to integrate a series of links, such as algorithms, expert experience, project deployment, interface opening, and so on, in order to complete a set of solutions, far more than a single point of technology can be "guaranteed."

In this era of excessive belief in technology, it is important to know what technology can do, and it may be more important to know what it cannot do.

To this end, Lei Feng.com 's AI Financial Review has planned a series of topics on "can AI solve the problem of financial rigid demand", using the same topic to interview interviewees from different backgrounds, products and enterprises with different customer groups, hoping to capture the spark of clashing views under different business and technological perceptions.

In the first three interviews, we interviewed Gu Lingyun, CEO of Bingjian Technology, CEO Huang Ling of Huian Jinke, and Li Huike, Executive Vice President of Pinti.

In the fourth article in this series, Wang Lei, executive vice president of Bangsheng Technology, tells an interesting story about his many years of experience in "AI Financial risk Control".

The following is Wang Lei's personal experience:

The difficult problem of "headache"

For a long time, banks have had a headache about "credit card cashing".

Credit card is essentially a kind of loan. Banks want loans to be used for consumption, rather than investing in stocks, real estate or even betting. However, in real life, many people do not use the money in accordance with the rules, thus giving rise to the "cash-out organization".

For banks, if the money in the credit card is not used for consumption, but for investment and gambling, on the one hand, it increases the risk that the money cannot be returned, and second, it violates the national loan policy.

Therefore, the bank will identify and monitor the cash-out behavior through some technical means.

Bangsheng Technology has received an order from a large bank, and they have an excellent team that has accumulated a long time and experience in the field of risk control, but they still hope to climb another tall building on this basis, so they come to us, hoping to introduce new ideas to solve problems through the AI capabilities of financial technology companies.

At the time, the bank was well able to identify which individuals had acted as loan frauds, but there was nothing it could do about the ever-changing and well-organized "professional" loan fraud gangs.

In the initial stage of the project, customers have high expectations of us, thinking that we will be able to find some connection and identify fraud gangs through more advanced technology and methods and through the analysis of individual loan fraud.

At that time, many institutions had not yet reached this level, and even the best financial institutions in the industry had put a lot of effort into identifying these fraudulent organizations very accurately. At the same time, we were in the early stages of our business and lacked experience, so we had no confidence to accomplish this task at that time.

Of course, whether or not to have confidence and whether to do it are two different things. We decided to accept the challenge.

Caught thousands of criminal gangs in one breath.

Sample is an important prerequisite for intelligent risk control to play a role.

If you want to identify fraud gangs through AI, you still have to start with high-quality samples.

The quality of the sample depends to a large extent on the experience of experts. In the field of risk control, expert experience is a scarcer and more important resource than AI capabilities. For example, "how wool is", what kind of behavior is wool wool, experts will use years of accumulated experience and rules to identify which behaviors are "wool wool", which belong to card theft, which are money laundering, which belong to application fraud, and so on.

The excellent experts in our team analyzed the bank's samples and found that the quality of their samples was not very good.

So the experts processed the samples through many years of experience in the field of risk control, and obtained some samples that we think are better. On this basis, we use machine learning modeling platform and association graph platform to identify suspicious individuals and groups.

At that time, several members of our team uncovered thousands of "credit card cash gangs", thousands of accounts, and hundreds of thousands of problematic credit cards in more than two months.

After that period of continuous exploration, we have a sense of understanding of the whole industry.

Before, we thought it was very difficult when we didn't start to do the work of "identifying cash-out gangs". But next time I am doing something else, I think it can be solved, and the train of thought is very clear.

In fact, because of the large number of banks involved in that project, we were also faced with great risks and pressures.

But we withstood the pressure, accumulated many characteristics and portraits of fraud gangs, and further optimized the model.

Later, the bank conducted an investigation through the clues we provided, determined that these people really belonged to cash-out gangs, reduced their credit card lines, and locked some credit cards, which worked very well.

It is also because of this experience that I have a more intuitive understanding of the application value of AI in the "risk control field", which makes me more confident in the application of AI in risk control scenarios.

Today, Bangsheng Technology has done a larger project for the state-owned bank, a comprehensive, bank-wide anti-fraud system that includes anti-fraud applications and electronic channel transactions.

The "identification cash gang" has become an integral part of the entire large-scale project and continues to provide services to the bank.

Seize the heart of the bank

Only by thinking about what you think and solving your worries can you catch a person's heart.

By the same token, if you want to get a big order from a bank, you must first know it, and then solve the problems it dreams of solving.

When dealing with bank customers, it may not be satisfied with all the time and all the products. At this point, we need to quantify the effectiveness of the project, immerse ourselves in its point of view, and make it trust you.

Therefore, we will always track the effectiveness of AI products and develop a series of visible and palpable indicators to give it a clear understanding of the effectiveness of our work and build a sense of trust.

For example, to intercept the number of illegal amounts, the degree of interference to customers, and the level of risk, we will record these data or results about risk control as a standard to count the final model good or bad.

We generally advise customers to adjust or optimize the model every three or six months. Frequent adjustment is not the style of the bank.

For banks and other financial institutions, "stability" trumps everything, followed by improvement. Therefore, they are also more cautious about the application of new technologies.

Banks will not easily change their existing risk control methods, because change means uncertain risks. Only when it is observed that the technology and effects are really stable will they adopt the new technology system.

AI mainly prevents and controls two kinds of risks in the credit field, one is fraud risk, the other is our common credit risk, such as "sesame credit" and "WeChat Pay score" that we young people are very familiar with.

When a loan is issued, the bank must first judge whether it is at risk of fraud, and then judge its credit risk.

First of all, the biggest worry for financial institutions is to be defrauded by swindlers and give loans to fraudsters.

When swindlers are shut out through anti-fraud technology, banks also have to worry about whether ordinary people can pay back money normally. Ordinary people may not pay due to factors such as poor business management, loss of jobs or changes in the social environment, so they have to judge their credit risk.

In the assessment of credit risk, the "scorecard model" is a very well-explained and stable model, and it is also widely used at this stage, and banks are used to using this traditional way. Therefore, there are not many scenarios in which we need the technical capabilities of our financial technology companies in this field.

In the field of anti-fraud, more and more financial institutions begin to accept machine learning model. Because fraud is more hidden, fraud risk is more difficult to control than credit risk.

And the industry generally takes six months to train the risk control model, and the online model takes at least three months, which adds up to nine months. During this period, the form of fraud in the whole market will change greatly, and when fraud gangs find that financial institutions will prevent and control fraud, they will change their tactics.

In view of this situation, Bangsheng Technology has specially set up a link in the model training-feature engineering. In this link, we try to show more features, we input thousands to tens of thousands of features into a model to make it cover more possibilities, although it can not completely solve the problem, but through such an algorithm to tune, the cost of fraud will be higher and higher.

In addition, banks need to consider whether the technology they use meets the requirements of regulatory policies.

One of the biggest problems with machine learning techniques is opacity and inexplicability. It uses a nonlinear algorithm, when the model says that the loan can be put, but its reasoning process is irreversible and can not be explained by words, so it will be subject to some restrictions in the process of use.

The model itself is the cooperation of an algorithm, and we are also trying to enhance the interpretation of the model through some technical ways.

AI is not a panacea

Since its inception, Bangsheng Technology has cooperated with many large banks, such as Agricultural Bank of China, China Construction Bank, China Merchants Bank and so on. Its understanding of the industry has also experienced a process from scratch.

But AI also has many problems that can not be solved.

Just mentioned that the effect of AI risk control depends on the sample. In scenarios such as fraud, samples are naturally sufficient, while in scenarios such as credit card theft and account embezzlement, because banks have stronger prevention and control efforts and fewer cases, fewer samples can be accumulated. the effect of intelligent risk control is also uneven.

Now, machine learning and AI applications are being tried in all areas of finance, but each bank pays different attention to the accumulation of samples, so some do better, some do generally.

At present, we rely too much on samples, so many companies in the industry are also trying unsupervised or semi-supervised technology.

Supervision means samples, no supervision means no samples, and semi-supervision means that the quality of samples is not high enough.

The supervised sample is that the team accumulates little by little from the business, inputs it into the model through the rule system and expert experience, and then the model runs.

In an unsupervised way, it is generally because the team does not accumulate samples and there are no experts in this field, so rely on pure algorithms to find out high-risk financial transactions through mathematical relations such as aggregation and clustering. Extract these abnormal samples and let the experts judge. The unsupervised process is to get the results through the model and let people judge, and then the model is optimized according to the results of human judgment.

In practice, no matter which technical method we use, we will advise customers to adopt a comprehensive solution instead of a single AI product, which includes expert experience, rule system, model system, graph system, big data calculation and so on. These elements must be combined to form a joint force to solve the problem.

Most of the problems in society themselves are very complex things, which need systematic thinking and technology to solve. We can not think that relying on a certain high and new technology can solve all problems.

We recruit a lot of new people every year, and we also come into contact with many new entrepreneurs. My overall feeling is that people are a little too superstitious in technology and high-end algorithms.

The application of any technology has prerequisites, such as data environment and sample quality. Each adjustment to the AI model, at least three months, more than a few years, the operating cost is also very high.

If the environmental conditions do not permit, it will be very difficult for the technology to achieve the desired state.

So for many of these new talents, I suggest that we first have a deep business understanding of this field and see if we have no ability to create a good environment for technology.

Without a deep understanding of the business, the data that does not meet expectations can not be transformed into high-quality samples, and the model is difficult to run.

Https://www.leiphone.com/news/202006/ahRf3GMGF2J7lNtI.html

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.