Intelligent risk management – is one of the most important application scenarios in the field of financial technology. With the help of big data and artificial intelligence technology to improve risk management capabilities, it is increasingly used in banking, securities, insurance, Internet finance and other fields.
In recent years, financial institutions like to mention the concept of smart risk control, but there is no accurate definition of smart risk control in the market. In fact, before the arrival of the AI wave, no one connected risk management with “intelligence.” After the big data fever has gradually cooled down, with the rise of artificial intelligence, the field of measurement finance has ushered in a new wave of hot spots: AI in Financial Risk Management.
I hate the buzzword “intelligence” because it is just a purposeful packaging of machine learning, and machine learning is far from being developed to a reliable engineering principle, and there are many problems of uncertainty, scale, and reasoning. As the use of big data and AI technology in risk management has become more and more common, financial institutions have an urgent need for the construction of intelligent risk control systems, and the market maturity and concentration have gradually increased. I want to focus on its connotation, application model and application architecture, and talk about what intelligent risk control is.
There are many kinds of “alias” for intelligent risk control. I have seen big data risk control, decision engine, risk measurement engine, risk model laboratory… in fact, they can all be included in this concept. Its basic logic is to use the computing and analysis capabilities of the big data platform, machine learning or deep learning models, and apply it to credit risk control, anti-fraud, anti-money laundering, transaction monitoring, insurance claims and other scenarios. So, it is essentially data-driven risk management and operational excellence.
For financial institutions, intelligent risk control does not change the underlying business logic, nor is it essentially different from traditional financial risk control models, modeling methods, and principles. It is only because of the introduction of big data that more dimensions of external data can be obtained, such as customer behavior, e-commerce consumption, operator data, geographic location, shopping habits, and so on. Compared with traditional financial data (such as central bank credit, transaction flow, asset status, financial statements, etc.), although these data are not necessarily related to customer defaults, they add more risk factors and variables, which can be viewed from more levels Describing the customer risk view can improve the effect of risk pricing and default calculation.
Intelligent risk control has changed the past risk management model oriented towards compliance and meeting regulatory inspections, emphasizing the use of financial technology to reduce risk management costs, improve customer experience, and data-driven risk control energy efficiency, which essentially represents a lean risk management thinking. If the comprehensive risk management of the New Basel Capital Accord (capital measurement, supervision and inspection, and market discipline as the three pillars) is the embodiment of traditional risk control, then intelligent risk control is the change and innovation of risk management practices in the Internet and big data era.
From the big data risk control in the past few years to the intelligent risk control today, the processing logic of data + model + rules has not been changed , but the application of machine learning models such as linear regression, logistic regression, support vector machines, and neural networks have been highlighted, and deep learning, integrated learning, etc. For the risk management personnel of financial institutions, with the increase of external data acquisition channels, the risk data has been effectively supplemented for people or small and medium-sized enterprises who were unable to obtain effective risk characteristics in the past, so that they can be designed for more scenarios and groups of people. Different financial products play the role of financial technology empowering inclusive finance.
The application model of smart risk control in the financial sector should be viewed from the perspective of different industries. Although they are essentially data-driven risk control and management decisions, due to the large differences in the industry attributes and business scenarios of banks, securities, and insurance, the application models of intelligent risk control are also different.
Banking industry: credit, anti-fraud, correlation analysis
Although it has not been verified, I believe that the title of intelligent risk control should originally be derived from the banking industry’s big data application in credit risk management, transaction anti-fraud, risk pricing, and monitoring of related relationships. Companies like FICO, Experian, and Equifax have already implemented various risk control models to achieve anti-fraud or credit reporting. With the enrichment of technical means and the gradual convenience of data acquisition, commercial banks can acquire, store, and process data of different dimensions through external data cooperation. Through the powerful computing power of the big data basic platform, calculate the correlation between users, such as phone number, mailbox, address, device number, etc.
Take consumer credit risk control as an example. According to the time dimension of risk control before, during and after loan, credit quality, solvency, collateral value, financial status, and repayment conditions are used as evaluation dimensions. Time and evaluation are formed Different credit risk focus points. Commercial banks combine different credit risk concerns to obtain relevant data.
In addition to big data, the “intelligence” of intelligent risk control is mainly reflected in the machine learning algorithm construction model. After business objectives such as credit application, default loss calculation, overdue prediction, anti-fraud, etc. are determined, use methods such as the integration of internal and external data, preprocessing (such as sampling, PCA, missing value filling, normalization), and feature statistics, and then select Appropriate algorithm for analysis. In the application of basic tools, due to the application of big data technology, it is mostly inseparable from basic computing platforms such as Hadoop/Spark and data analysis tools such as R/Python. At present, some manufacturers that focus on machine learning, such as Fourth Paradigm and Alibaba Cloud, have also developed draggable modeling tools, which partially reduce the learning cost and threshold of machine learning.
Whether it is bank loans, mortgage or guarantee loans to individuals or companies, supply chain loans, score cards, loans in the Basel Agreement, or the current popular smart risk control, the fundamental principle is to measure customer repayment ability and willingness. Intelligent risk control only uses more data dimensions to describe customer characteristics, so as to more accurately quantify customer default costs and achieve reasonable credit to customers. It can be seen that its principles and methodology are no different from traditional financial risk control, but manual review can be replaced by automated approval to reduce labor costs.
Securities industry: abnormal trading behavior, illegal account detection
Different from the intelligent risk control of the banking industry which focuses on credit risk control and anti-fraud, securities companies and exchanges pay more attention to the detection of “real-time” and “in-the-event” transaction violations. In terms of regulatory requirements, the Shanghai and Shenzhen Stock Exchange also recently issued the “Notice on Strengthening the Management of Key Monitoring Accounts”, which requires strengthening the front-line supervision of transactions, highlighting the supervision in the event, and clarifying the monitoring of key monitoring accounts for serious abnormal trading behavior; from technology On the one hand, due to the large amount of data and high concurrency during the daily intraday continuous trading phase, low-latency real-time calculations, machine learning, and complex event processing are the key points of the design of risk control for securities intelligent trading.
The characterization of abnormal trading behavior is essentially a user portrait project, which divides high-frequency trading customers into groups, establishes a user portrait system, extracts features based on various indicators in customer transaction behavior, uses these features as the input of the model, and the output is the category to which the user belongs. There are characteristic indicators such as transaction activity (number of orders, frequency of orders, etc.), quotation of each order, holding target, total assets, funds and position information, etc. At the level of securities business, it needs to cover brokerage business, self-operation, asset management and other businesses.
For the current lively smart risk control market, I think its problems can be summarized into three points:
Firstly, many so-called smart risk control systems on the market still have nothing to say about “smartness”. Most of them are still based on the combination of rules and conditional screening to achieve risk warning, and semi-automatic methods to assist people to judge. Some systems actually make a customer portrait and labeling system, while others even package risk indicators and report systems into intelligent risk control platforms. There is no standard in the industry, and there is no lack of deliberate exaggeration and exaggeration.
Secondly, for financial institutions, the prerequisite for intelligent risk control is the use of big data, and big data is completely dependent on the layout of the scene. Internet companies are naturally able to accumulate massive amounts of data to characterize user behavior due to their stable scenario advantages, while most financial institutions still need to resort to external data transactions to supplement the lack of data dimensions. On the one hand, the purchased data may lack financial attributes and have no reference significance; on the other hand, because financial institutions lack a stable and sustainable scenario and data operation system like the Internet, it is difficult to form a real role and value. Therefore, some institutions simply buy Tongdun points and Zhima credit points directly instead of building the system themselves.
Thirdly, big data emphasizes relevance rather than causality, and the evaluation and results based on machine learning models are not interpretable. The feasibility of applying it to actual risk review and detection scenarios is open to question.
Therefore, intelligent risk control can only be regarded as an emerging technology application model in the field of risk management in the current big data and artificial intelligence outlets. In terms of risk management methodology, risk management measurement standards and regulatory requirements, they are no different from traditional financial risk management models. Its characteristic lies in the introduction of more dimensional customer data (such as customer online consumption, e-commerce, operators, behavior), reconstruction of data and application architecture, and the use of distributed big data platform capabilities, machine learning or deep learning models , To achieve large-scale, fast and accurate risk event filtering or prediction, so as to be able to seek breakthroughs in the timeliness, forward-looking, precise, and technological advancement of risk management. In the future, as biometrics, image recognition, blockchain and other technologies mature, they may be used in the field of risk management to form more intelligent risk control application models.