When it comes to the concept of risk in relation to financial sector, first of all, the retail lending segment comes to mind for many. And we are talking, accordingly, about credit risk. At the same time, in lending, risk has long been learned not only to assess, but also to manage it. Credit risk calculated using predictive methods for assessing the likelihood of a borrower's default in the future. Specially developed scoring models have been successfully coping with this for many years.

As for insurance, even for some players in this market it is still a surprise that it is possible to determine the risk of unprofitability of a policy by analogy with calculating credit risk. That is, using the same scoring. The loss ratio of the policy, that is, the ratio of payments for insurance events to the collected premium, is the target risk variable in insurance and, at first glance, has nothing to do with loan default. But in fact, both of these events have a common nature - lack of accuracy and neglect of their own obligations on the part of the subject.

With the introduction of amendments to the law “On Credit Histories” a year and a half ago, insurance companies now have the opportunity to obtain the credit histories of their clients. Since for insurers, as well as for lenders, credit histories are of the greatest interest precisely from the point of view of the possibility of assessing risk, the insurance industry and, accordingly, the NBKI (which stores the credit histories of 74 million Russians) faced the question of constructing a mathematical model, predicting unprofitability based on data from credit histories - insurance scoring.

This dependence has long been discovered and actively used by insurers. different countries. In Russia, such a correlation was known before, but could not be used in practice until 2014: the Law “On Credit Histories” did not allow credit history to be provided to non-creditors. Almost immediately after the amendments came into force, work began to formalize the mentioned dependence. The work was attended by NBKI experts, actuaries from major insurance companies and specialists from FICO, the author of the most popular and effective insurance scoring in the world.

By mid-2015, more than 5 million insurance policies had been processed and the match with the credit history database was about 80%. Insurance scoring, calculated on the basis of credit histories, as in retail lending, takes into account the quality of servicing of loan obligations, types of loans and history of use of borrowed funds. For ease of use, NBCH and FICO retained the scoring model scale - from 350 to 850 points. A low score means a high risk of unprofitability of the policy, and a high score means the opposite.

The results of testing the model on real auto insurance policies turned out to be comparable to credit scoring: CASCO, for which the model calculated a low scoring score (less than 625), turned out to be 20% unprofitable than policies with a high score (more than 725). This result was confirmed both for Moscow policies and for regional ones. Even more clear results were obtained when analyzing loss rates from specific insurance events. For example, for damage from car theft, the loss rate of policies in low scoring ranges is 5 times higher than for the upper range. Obviously, this is due to the fact that the NBKI insurance scoring was able to identify unscrupulous citizens to whom banks had already stopped lending money due to their poor payment discipline and high debt load, and they went to insurance companies, hoping to solve the problem with the help of insurance payments and deception. their financial difficulties. In other words, NBKI insurance scoring has proven useful in preventing insurance fraud.

And finally, the success of the work done in auto insurance allows us to hope that similar technologies will be applicable in other types of insurance. According to NBKI and large insurance companies, the search and validation of dependencies between a person’s responsibility and his behavior for most insurance products is a matter of the near future.

In what types of insurance is scoring relevant?

So far, only credit history bureau scoring has received widespread use in Russia - in motor insurance. The identified dependencies make it possible to significantly clarify the forecast of loss under a comprehensive insurance policy and even counteract attempts at property fraud. For example, the comprehensive insurance tariff depends on the age, gender, marital status of the car owner, the brand and region of operation of the car, as well as other parameters that insurers call the tariff factor. According to the deputy general director, Risk Director - Head of the Department of Actuarial Calculations at Sberbank Insurance Vladimir Novikov, this is scoring. With the development of digital technologies and the accumulation of large volumes of data, it has become possible, in addition to the classic risk assessment factors, to use those that previously did not attract the attention of underwriters. The scoring technique is applicable not only to risk assessment: it works well in solving problems of marketing, sales, optimization of claims settlement, and combating fraud, believes Vladimir Novikov.

According to Head of the Marketing Research Department of IC "MAKS" Evgeniy Popkov, in the recent past, insurance scoring was a very limited toolkit. Thus, in most cases, sales office employees used insurance calculators for voluntary types, in which the control was triggered by certain triggers - “Underwriter approval required” or “Security check required”.

Alexander Morozov, Director of Statistics and Analytics, Smart Driving Laboratory, states that scoring is essentially a personal assessment of insurance risk. This estimate is more accurate compared to traditional models calculated based on averaged factors.

Alexey Danilov, CEO of Adaperio, gives the following example. Traditional methods assessments have always been based on the behavior of the average user - an abstract policyholder of a certain socio-demographic profile, but in fact the behavior of, for example, two 35-year-old men living in Moscow and using a BMW can be radically different. It is in this case that big data becomes useful, which will make it possible to more accurately determine the risks of the insurance company and, as a result, affect profit (loss) indicators.

How to learn to detect auto insurance fraud using machine learning methods? About this using the example of a scoring model with lift equal to 4. Ilya Lopatinsky, director of support department retail business Ingosstrakh will tell you at Scoring Days 2018.

In world practice, scoring is used in all lines of business of insurance companies. IN Russian practice scoring is most common in types such as voluntary health insurance and car insurance, says General Director of BKI Equifax Oleg Lagutkin.“The most exotic type of application of scoring in our practice was assessing the propensity to fraud of insurance company employees who make decisions on the terms of concluding an insurance contract,” says Oleg Lagutkin. In his opinion, it is advisable to introduce scoring into processes such as antifraud, losses and sales.

Deputy Director of the Underwriting and Product Management Department of Soglasie Insurance Company Andrey Kovalev sees the potential for using scoring in all voluntary mass types of insurance (including auto insurance, voluntary health insurance, personal income insurance). The main area of ​​use of scoring is risk assessment and anti-fraud, but it can also find application in the field of sales support.

Deputy General Director of VTB Insurance Evgeniy Nisselson believes that scoring is more appropriate to use in sales of retail products, such as auto insurance, property insurance, accident insurance, etc. It allows you to reduce the cost of risk assessment and significantly speed up this process. Scoring is applicable to standard products; to analyze specific risks, it is necessary to use traditional methods.

Maria Barsova, Operations Director - Deputy General Director for Property Types of Insurance of SAO ERGO, said that the company uses credit scoring in comprehensive insurance and individual insurance individuals, mainly in underwriting and for determining tariffs.

Insurers are testing telematics

According to Dmitry Rykov, head of the underwriting department at auto insurance LLC Zetta Insurance, telematics-based policies have not yet received large-scale development, but the company continues to carefully test these products, monitor the market and is preparing to make an interesting offer. IN SC "Soglasie" also confirmed that the implementation of scoring on these telematics devices is in the development and testing stage. IN « VTB Insurance» reported that the insurer does not use scoring based on telematics data on an industrial basis due to its limited presence in the auto insurance market. At the same time, the company tested telematics systems from different manufacturers and the results showed quite high efficiency. Maria Barsova, Operations Director - Deputy General Director for Property Types of Insurance of SAO ERGO, said that the company introduced scoring based on data from telematic devices and continues to do so, but it cannot be said that expectations were met 100%. The volumes are still small, and therefore it is too early to talk about the impact on unprofitability.

“Any data is useful for improving the assessment of personal insurance risk. Moreover, if they correlate well with this very risk and have no analogues. Data from telematics devices received directly from the car cannot be qualitatively replaced by other factors and correlates well with insurance risk, noted Alexander Morozov, Director of Statistics and Analytics, Smart Driving Laboratory. – Therefore, we can say for sure that telematics data is useful for scoring. The result of implementation depends on the specific model proposed by the insurance company, the composition, quality and cost of the data itself, so it would be incorrect to name any single assessment.”

What data should be used when constructing the scoring for comprehensive insurance? About this in the speech Frank Shikhaliev, head of the data analysis development department at Renaissance Insurance on April 19 at Scoring Days 2018.

Technologies: what insurers use

When asked whether the company uses its own developments or those of third-party suppliers, the company "Agreement" stated that they use both of these approaches. “Undoubtedly, with internal developments, higher business sustainability is ensured, but there are still areas where the company cannot carry out all developments on its own,” said Andrey Kovalev, Deputy Director of the Underwriting and Product Management Department of Soglasie Insurance Company. VTB Insurance Company uses ready-made solutions from suppliers, customized to the needs of the insurer. Company work "Sberbank insurance" within the framework of scoring, it can be divided into two parts. One part is the analysis, where software and statistical packages that were developed for the company by third party contractors. The second part - the remaining 50% of success in using scoring data - is determined by the competence of employees, that is, it depends on the presence in the company of specialists who can work with big data.

Head of underwriting department in auto insurance of Zetta Insurance LLC Dmitry Rykov said that in addition to its own methods, the company uses tools provided by partners. One example is the Audatex service, which allows you to check the accident history of a car. Another example is the CBM for compulsory motor liability insurance, which also allows you to roughly assess the client’s insurance history.

Scoring insights from insurers and developers

The choice of car model really carries information about the client’s behavior on the road. For example, a client who has chosen a brand of vehicle that emphasizes the driving properties of cars consistently gets into accidents more often than a client who chooses a vehicle of the same class, power, size and cost, but from a manufacturer that emphasizes comfort or reliability, said Andrey Kovalev, Deputy Director of the Underwriting and Product Management Department of Soglasie Insurance Company.

Cases of scoring in car insurance from Ilya Lopatinsky from Ingosstrakh and Frank Shikhaliev from Renaissance Insurance - at the conference Scoring Days 2018.

According to Dmitry Rykov, head of the underwriting department at auto insurance LLC Zetta Insurance, there are many interesting dependencies: for example, the frequency of accidents for policyholders in different family statuses varies significantly. Thus, married drivers have the lowest frequency of insurance claims and receive a discount from the company. Another relationship that the company discovered directly in Moscow is the relationship between the probability of occurrence insured event and permanent registration addresses of the policyholder. The discount for a car owner living in an area with safer traffic can be 20% of the policy cost.

Vladimir Shikin, Deputy Director of Marketing at NBKI, reported that, as a rule, all patterns have a logical explanation, but it happens that they are discovered after the fact. For example, during testing, the company noticed that in a segment with low bank scoring values, there is a high probability of loss from theft. “We made the assumption that in this range there may be clients to whom, due to low responsibility, banks no longer give loans, and these people can solve their financial problems at the expense of insurance companies. That is, in essence, we have identified an indicator of potential fraud,” said Vladimir Shikin.

In the summer of 2014, as a result of amendments to Federal Law 218 “On Credit Histories,” insurance companies were able to obtain the credit histories of their clients. As for lenders, for insurance companies, credit histories are of maximum interest from the point of view of the possibility of assessing risk. In lending, risk assessment is necessary to predict default; in insurance, to predict the loss ratio of a policy. Thus, the insurance industry and the National Bureau of Credit Histories (NBKI), the place where the credit histories of 72 million Russians are stored, were faced with the question of building a mathematical model that predicts unprofitability based on data from credit histories - insurance scoring.

Large Russian insurance companies, NBKI and the international leader in the field of predictive analytics, the FICO company, took part in the creation of insurance scoring based on credit histories. FICO's experience in the American and European insurance markets was taken as a basis research work and largely ensured obtaining a quick and strong result from a mathematical point of view. Thanks to international experience We immediately began to focus on motor insurance. This industry worldwide shows a strong correlation between insurance policy loss rates and customer personal liability.

To begin the research exercise, a hypothesis was formed about the most powerful predictive variables from the credit history. At this stage, the experience of building a credit scoring system that predicts the borrower’s default was used. credit obligations. As in insurance, credit process the lender evaluates the client’s responsibility – his personal characteristics, built on the basis of the history of fulfillment of previously assumed obligations. Each variable underwent a thorough analysis on the generated database of historical insurance policies; as a result, the strongest and most stable variables were selected.

Among the most powerful variables is, of course, data on breach of obligations. The number and depth of overdue payments have a downward impact on the scoring score. On the other hand, experience in using long-term loans is welcome: positive experience with mortgages and car loans has an increasing impact on the result. One of the most complex issues To build a scoring model, regional specifics were taken into account. As a result, the final model included several variables based on regional data.

Minor variables, but nevertheless influencing the scoring score, include data on the client’s family members. This information is included in the scoring model in order to take into account the situation in which, for example, one of the spouses takes upon himself all issues of interaction with creditors, although the family has a common economy. That is, relatively speaking, a woman’s ideal credit history does not mean at all that she will not have problems if her husband has continuous violations of his obligations.

Validation of variables and the scoring model as a whole is a daunting task. Its successful solution largely depends on the representativeness and volume of data available for retrotesting. In this regard, it is necessary to pay tribute to the leaders Russian market auto insurance. All of them were involved in the creation and validation of the model literally from the first days. The numbers speak for themselves: the total number of comprehensive insurance policies participating in the analysis is more than 6.5 million. The historical retrospective of the policies is more than six years. This made it possible to create a model separately for Moscow and the regions and ensure its stability - measurements were carried out over five reference periods with an interval of one year. Not only was the correlation of the scoring score with the overall loss ratio for policies studied, but also the dependence was built on individual types - for example, the loss rate from theft.

As a result, the resulting scoring model showed: the loss rate for policies in the lowest scoring range of up to 600 points (the insurance scoring scale is adjusted to most popular FICO scoring systems: from 350 to 850 points, with lower score values ​​meaning greater risk) is on average 20% higher, than in the range of 700 points for Moscow and 30% for the regions.

Interesting results were obtained by studying the correlation of the scoring score with the loss ratio for certain types of compensation. For example, when studying the dependence of payments for car thefts on scoring points, an anomaly was identified - a sharp increase in unprofitability (4-5 times) in the range below 550 points. Consultations with colleagues made it possible to provide an explanation for this phenomenon: citizens with low payment discipline and excessive debt burden can no longer receive borrowed funds from creditors, because they are denied, and are trying to solve their financial problems at the expense of insurance companies. That is, in fact, we are talking about insurance fraud. As it turned out, insurance scoring based on credit histories can effectively counter this threat.

The results obtained open up prospects for using insurance scoring based on credit histories in motor insurance in the short term. Firstly, insurance companies can already use NBKI scoring to set comprehensive insurance pricing and make decisions about selling policies to specific clients. For example, using increasing coefficients for high-risk segments. Secondly, insurance scoring is applicable to predict the loss ratio of a portfolio - an extremely important task that actuaries regularly face and the accuracy of which largely depends on financial results the entire company.

And finally, the success of creating a scoring model in auto insurance allows us to hope that similar technologies will be applicable in other types of insurance. According to NBKI and large insurance companies, the search and validation of dependencies between a person’s responsibility and his behavior for most insurance products is a matter of the near future.

Dedicated to the possibilities of scoring in the financial sector. Bankers and microfinance organizations shared successful cases, and IT companies and mobile operators talked about new opportunities. Unfortunately, there was not a single representative of the insurance industry among the speakers at the conference.

Is scoring, as an analysis tool, really not interesting to the insurance community? Quite the opposite. But if banks have long mastered this technology for analyzing the client base and widely use it when lending, then the insurance market is not yet so spoiled by this method of client selection. Nevertheless, to one degree or another, insurance companies still turn to this tool to create more adequate underwriting.

Five years ago, insurers did not use scoring tools at all. Three years ago they began timidly trying to use credit scoring in connection with the “engine”. Today, credit scoring can already serve as one of the key metrics in auto insurance underwriting and is gradually used in working with other types of insurance.

We, just like banks, want to know our clients in person. In order to correctly form a reserve and set a tariff, it is very important to understand what kind of person is in front of you, what can be expected from him, how unprofitable a particular client can be. Numerous studies carried out on the market financial institutions, have already proven that if a person is undisciplined in one area of ​​life, then he is likely to be undisciplined in other areas. Financial discipline, behavioral patterns and habits - this is what banks have been interested in for a long time, and now insurers should rightfully be interested as well.

There are a lot of data sources for collecting information: from credit bureaus to social networks, which can tell a lot about the client. The choice of these sources is determined by the specific needs of the company, budget and functionality of IT systems.

But the most important question is not what data to analyze (there are indeed many options and sources now), but how to do it. It is the correct interpretation of data, placement of emphasis and weights that allows you to build a working scoring system, which will not only help you understand the potential unprofitability of a particular client, but will also allow you to identify fraudsters who can lead the company to serious financial losses.

According to FICO and NBKI, which are actively conquering the niche of insurance scoring based on credit history data, clients with a comprehensive insurance policy with a low scoring score show a loss rate several tens of percent higher than those with a high scoring score. Having such data, how much Insurance Company will be able to reduce the unprofitability of the portfolio? It is difficult to give a definite answer to this.

This indicator largely depends on the insurance segment and especially on how exactly to use the scoring result (refuse insurance altogether, offer an increasing coefficient, or something else). In marginal forms, it can reach several percent, and if a company’s portfolio amounts to billions of rubles, then the benefit can amount to several tens of millions.

The second difficulty is cost. Despite the fact that over the past few years the price of analysis of one client has decreased almost threefold (different data operators have different prices), scoring is still used mainly only in car insurance. Thanks to high margins, this is where the additional costs of analyzing the customer base are justified. For other types (property or accident insurance), scoring is still used more as part of experiments rather than for real savings.

The justification for scoring costs is also related to the volume of the analyzed portfolio. In our country there is quite high level debt burden of the population, the volume of loans continues to grow, even despite the fall in real incomes. At the same time, the penetration of insurance services is extremely low. Only this year we began to gradually increase the share of penetration in property insurance and life insurance. But this, of course, is not enough.

If the market manages to overcome at least one of these obstacles, then scoring in insurance will most likely cease to be almost a fantasy, becoming an effective stage of high-quality underwriting. After all, the potential of this tool is really very high.