US zestfinance: Financial risk control with large data

Wen/People's Bank of China Credit Center postdoctoral Liu Nao Hai, Turbo Financial Group co-founder, chief risk officer Gu Lingyun, China Unicom Network Technology Research Institute, senior engineer Ding

As an innovative technology financial company, ZestFinance has been the focus of investment and Internet finance since its inception, because of its unique business philosophy. In 2013, the global third-party payment platform PayPal co-founder, the United States well-known investors Thiel (Peter Thiel) 20 million U.S. dollars in investment. ZestFinance that its mission is to create fair and transparent credit information for everyone. ZestFinance's initial service was to use usury groups, known as loan-day lenders, to dig out their credit information through large data to help them enjoy normal financial services. ZestFinance that every consumer is a "good" person, hoping to help them realize their right to a normal financial service by gathering evidence to prove the true credit status of the people with incomplete credit information.

As with the traditional levy, ZestFinance's credit rating for consumers is based on two fundamental information: the ability of consumers to repay and the willingness of consumers to repay. The difference is that the traditional credit data relies on bank credit data, and the data of large data collection not only includes traditional data, but also some descriptive risk characteristics related to the repayment ability and repayment intention of consumers, These correlations describe the extraction and selection of risk characteristics as the core of zestfinance (Fig.). Compared with the strong correlation of traditional credit data, the data of these data is relatively weaker than that of the consumer, ZestFinance uses large data technology to collect more data dimensions to strengthen the descriptive ability of these weak related data. In this way, the large data credit is not dependent on the traditional loan data, it can be used to carry out the credit to the people who are unable to serve the traditional letter, and realize the coverage of the whole consumer.

A comparison between the view of data credit and traditional credit

Large data mining technology superior to bank

The advantage of Zestfiance is its strong ability of data mining, which can develop a novel credit evaluation model and excavate more difficult to find consumer credit information from large data.

Different applications have different understanding of large data, and there is no unified concept at present. ZestFinance's founder and CEO Merrill (Douglas C.merill) believes that "more data" is not "big data", and can use it technology to integrate fragmented information to form truly useful large data. This understanding of large data is particularly useful for credit-card industry, because the basic process of credit is also the integration of local information, which is dispersed in the unlikely usefulness of different lending institutions, into global information that can fully describe the consumer's credit status.

It is worth pointing out that although ZestFinance uses large data technology to carry out the credit, it mainly utilizes large structured data that is less utilized for complex large data types, such as text and social network data, The main reason is that these complex big data and the ZestFinance service consumer's credit risk correlation is too weak. This phenomenon is also verified by other internet banking practices, such as the largest Internet peer-to-peer company, Lending Club, which first landed in the credit market from Facebook, hoping to pass credit for the social networking data, which proved ineffective and unsustainable, Finally had to return to the traditional credit means development.

ZestFinance's core competence lies in its strong ability of data mining and model development, and it creatively applies to the traditional credit risk management field with the more mature technology in machine learning field.

The key is multidimensional data and algorithms

Zestfinane's core business is consumer credit approval, the main customers are subprime consumers, the main competitors are banks or pawn shops. ZestFinance's core competence lies in its strong ability of data mining and model development, and it creatively applies to the traditional credit risk management field with the more mature technology in machine learning field.

Traditional credit scoring models usually have 500 data items, from which 50 variables are extracted, and a predictive analysis model is used to make a quantitative evaluation of credit risk. In the new model of zestfinance, 3,500 data items are used, from which 70,000 variables are extracted, and 10 predictive analysis models are used for integrated learning or multi-angle learning to obtain the final consumer credit score.

As shown in Figure II, the ZestFinance data source is a large data, can generate tens of thousands of risk variables, and then input different predictive models, such as fraud model, authentication model, prepaid capability model, repayment capacity model, repayment intention model and stability model. Each model predicts the credit status of individual consumers from different angles, overcomes the limitation of a model consideration in traditional credit evaluation, and makes the prediction more meticulous.

Machine learning methods are widely used in production, research and life, while integrated learning is the most popular research direction in machine learning. Integrated learning is a machine learning method that uses a series of algorithm models to analyze and predict, and uses some rules to integrate the results of each model to obtain a better predictive effect than a single algorithm model.

If a single model is compared to a decision maker, the integrated learning approach is equivalent to multiple decision-makers making a decision together. Due to the integration of multiple information and multiple decision-making mechanisms, the analysis and prediction of integrated learning is better than that of a single model. The information in different angles is related to each other, which contains complementary information and multi-angle learning process, which is equivalent to a process of collecting evidence, strengthening complementary information and merging information. For example, the two independent scoring models are 16.9% and 9.4% of the increase in profitability, and the second model may be discarded in traditional credit evaluation, but if the two models are found to contain complementary information, the results of the two models can be fused to increase the profit to 38%.

Each of these multiple-angle learning models do not use traditional logic regression, but other predictive models in machine learning (the details of the model are core secrets for zestfinance). The reason why the logistic regression model is not used in the credit evaluation of ZestFinance is that the data of large data collection is relatively fragmented, and the variables are too many, and the distribution of risk variables cannot satisfy the normal distribution.

Seizing credit approval management can control 80% of the risk

According to the survey, about 80% of credit risk from the credit approval link, once the consumer access to credit, follow-up management can only control 20% of the risk, this shows that scientific credit approval management is very important. To develop high quality credit examination and approval scoring model and carry out scientific examination and approval risk management can greatly reduce the bad debt rate and obtain better economic benefit. Although ZestFinance did not disclose its bad debt rate, in the actual application process, compared with its rival banks or pawn lenders, they have achieved some good results.

1. The cost of obtaining a loan customer is 25% of that of the competitor. According to JMP2012 's industry report, for an online borrower, the average cost of acquiring a customer should be between 250 and 500 dollars. 2014, through ZestFinance services, the cost of obtaining customers is stable at around 100 dollars.

2. Default rates for first pay defaults, FPD, are below competitors. Since the beginning of 2012, ZestFinance's first loan default rate is still fluctuating, sometimes higher than the third competitor, with the continuous improvement of the model, by 2013 has been basically stable, significantly lower than three competitors.

3. Continuously improve the customer's ROI. ZestFinance's initial return on investment was around 100%, and as the model continued to improve, the current rate of return on customer investment reached more than 150%.

4. The return on investment is higher than that of competitors. Figure II shows a comparison of ZestFinance's investment returns with its rivals at different times, which shows that ZestFinance's ROI is significantly higher than industry standards and its rivals.

Figure II ZestFinance ROI higher than competitors

On zestfinance Credit evaluation model

The advantage of Zestfiance is its strong ability of data mining, developing a novel credit evaluation model (see figure III), and digging out more difficult to find consumer credit information from large data.

The scoring model is constantly being updated as ZestFinance continues to collect data and add new data sources. As shown in table one, a new credit evaluation model is introduced in almost every quarter from 2012 to the present. And the model is named after every different developer, and there are now 14 models. The improvement of ZestFinance scoring model also improved the level of credit risk assessment. While these new models still face the challenge of data adequacy and data availability, continuous improvement of the model is ongoing.

An interpretation of the credit evaluation model of Figure three ZestFinance

ZestFinance first engaged in credit approval, only credit rating model, and then continuously refine its evaluation model to support the continuous introduction of new credit risk business. As shown in Table I, a collection score was introduced in the first quarter of 2013 and a marketing score was introduced in the second quarter of 2014; Car loans and legal reminders were launched in 2014, and eight credit evaluation models have been developed for different credit risk assessment services.

Table A zestfinance credit evaluation model

The loss of data (Missing) refers to some data items in machine learning that are missing due to a variety of reasons, and more loss of data will pose challenges to the modeling process. Because the zestfinance uses the large data of multi-dimensional degree, the phenomenon of losing data is more prominent. ZestFinance processing of lost data has its own uniqueness. First, ZestFinance constantly improves its scoring model to enhance its ability to process lost data, and the latest scoring model can handle more than 30% of lost data. Secondly, zestfinance to explore the causes of data loss by making full use of the correlation between lost data and the intersection of normal data. Through such plowing, some useful consumer credit information is obtained. This, of course, requires the use of a particular environment, a combination of credit operations, and a deep understanding of consumer behavior patterns.

Revelation of large data credit to China's credit industry

ZestFinance's initial service was to use usury groups, known as loan-day lenders, to dig out their credit information through large data to help them enjoy normal financial services. ZestFinance that every consumer is a "good" person, hoping to help them realize their right to a normal financial service by gathering evidence to prove the true credit status of the people with incomplete credit information.

In contrast, the current domestic credit risk management, the punishment is too strong, a similar "guilt reasoning," the idea of a great way, this approach may be simple and effective, but did not give full play to the role of the credit. The true role of credit is not only to punish the breach of faith, but more importantly praise integrity. ZestFinance for all consumers to tap credit, with the power of science and technology to promote the development of Pu-hui finance, break the credit institutions for the rich circle of services.

Targeting specific service groups is also the key to zestfinance success. The main clients of ZestFinance are about 5% of the population, and the credit score is below 500. Through the in-depth understanding of this part of consumers, the screening of large data description information, developed is also targeted at this group of effective credit approval model. There is no special general analysis model in the field of machine learning, but it often has an effective model for a specific range. Therefore, it is the key to develop the credit analysis model and even carry out the credit service business to properly locate the service group and understand the service object.

In addition, an important reason why ZestFinance is superior to its competitors and traditional credit institutions is the ability to develop a strong credit scoring model: A model based on multiple-angle learning, which is updated and refined in a timely manner. In contrast, China's credit approval, or credit risk management, the level of uneven, from qualitative judgment to simple quantitative decision-making, overall quantitative analysis is insufficient, and credit rating agencies have not yet launched. Only by strengthening the investment of quantitative credit risk analysis technology, can we realize the professional risk management of consumer credit. At the same time, it is worth emphasizing that the big data age, there is no ready-made free lunch, data and models need to be refined, need data scientists to participate in the human, even if the zestfinance model to China, can not be used directly. The understanding of data and consumers and the mastery of data mining techniques are lessons that cannot be omitted from the modeling process.

(This article is only a personal opinion of the representative, regardless of the unit.) ）

(Responsible editor: Mengyishan)