Диссертация (Рандомизированные алгоритмы на основе интервальных узорных структур), страница 2

The Global Research InstituteMcKinsey estimated the total volume of information, stored by banks and financialinstitutions, in 1 exabyte (1018 bytes) in 2009 [3]. For example, Singapore DBS Bankis developing a project of using the IBM Watson supercomputer and is going to spendabout $11.8 million for these purposes within next three years.With the increase in data volume it becomes more difficult to find consistentpatterns in data, in particular due to "curse of dimensionality" problem. Traditionalmethods such as credit scoring, which were developed back in the 1950s to solve theproblems of risk management, do not always give satisfactory results. For example,applying logistic regression models to detect fraudulent transactions, has not shownappropriate results.

Banks and financial institutions have experimented, therefore,with more complicated training methods such as SVM, random forests, XGBoost[11].As far as history of statistical models in banking is concerned the first system ofcredit scoring appeared in American banks during the Second World War. In orderto set off a loss of the mobilized specialists, many credit organizations asked themto develop general decision rules of loan granting, which would be able to guideordinary employees.6The next stage in the development of scoring it was the appearance of consultingcompany Fair Issac Corporation in the early 50’s that specialized in the developmentof scoring models.

Today, FICO continues to dominate the development of scoringsystems. For instance, FICO scale which assesses quality of the borrower in the rangeof 400 to 900 points is wide spread and adopted by many credit organizations.Two additional factors had influence on further growth of credit scoring popularity: 1. Combating racial discrimination. The law on equal opportunities of receivingcredit, adopted in 1974, had obliged banks to consider loan applications with equalarguments [13]. 2.

The rapid development of computational capacities, which enabled credit bureau to process and store historical data of decent volume. As a result,scoring has taken the form of statistical model which allowed one to estimate probability of loan repayment based on credit history of known existing clients. Theprincipal limitation of this approach was quite obvious. The sample was labeled onlyamong clients who had already been given the loan.

It remained unknown how thoseclients would have perform if they were not declined with their application.The application of advanced machine learning methods in banks began withincredit risk management. And this is not accidental. Mistakes in loan granting accounted for the greatest loss in balance sheets. While in marketing the price formistake is foregone earnings then in a type of interest payments, then the mistake inmaking decision of loan granting is the all sum of debt, and not just percent.Fair Isaac Corporation was the first company that started providing credit assessment services of borrowers based on data analysis in 1958 [5].

Company actedas a consultant of commercial banks and sold scoring cards.In 1972 FICO introduced the first automatic scoring system in the Americanbank Wells Fargo. The first scoring cards were obtained with using logistic regressions. FICO Score is the development of the leader of modern scoring analytics,the American company Fair Isaac Corporation. Today FICO scoring systems havebeen fully deployed in 22 countries. Nowadays it is the most popular system on themarket, which allows to significantly reduce the credit risks of banks. The officialwebsite of the company claims that in the latest version of FICO8, the forecast accuracy is improved for 15The additions were made to improve the credit rating of7certain categories of clients in scoring model.

First of all, these were clients who hadrecently been actively engaged in the search of the new loan, as well as borrowerswith dark stains in credit history.Scoring model FICO Score is based on an assessment of the influence of 5groups of factors:1. The quality of credit history (timeliness of payments on loans). The importanceof the group in assessing of the individual rating is 352. The level of current debt of the client in the total amount of the credit limit.The importance of the group is 303. Quantitative analysis of credit history. For example, the duration of cooperationwith credit institutions.4.

The ratio of submitted loan applications and negative decisions made accordingto them.5. Analysis of types of early loans.The overall significance of groups 3, 4, 5 is 35%. The total amount of pointsscored generally in the range from 300 to 850.

If the individual rating of the borrowerwas lower 650 points, he/she is offered a loan under the conditions above market. Ifclient scored more than 720 points, he/she can count on sufficiently democratic creditconditions.First of all, under credit conditions are meant as an interest rate. The definitionfor high risk client profile associated with the decision problem called Risk BasedPricing. The differentiated principle of pricing of the general conditions succeeds forall categories of borrowers.Nowadays in credit programs of banks it is possible to see even more often therange of interest rates or the minimum rate to which then various extra charges areadded.

The final level of an interest rate is established after the analysis of the packageof documents provided by the borrower. In other words, the bank studies informa-8tion about the potential borrower, determines its category by bank risks, predicts itsbehavior and only then sets the interest rate for a loan.This technology is the pricing of risk or Risk-Based Pricing (RBP). In the USAand Western Europe it has become a norm long ago: if the borrower has positivecredit history, he/she can count on the minimum credit rate.

If, however, there isnegative information in the credit history bureau, the rate on the loan will be significantly higher. Today the personified credit conditions are offered also in manyRussian banks.Whether Risk-Based Pricing get accustomed in Russian reality? According toskeptics, the main obstacle for such credit programs is a high cost of technology.A bank has to have considerable resources and appropriate infrastructure, and onlylarge banks are able to afford it.

Another weak link is the Russian credit historybureau, which absolutely does not work very long at full strength. The request forcredit history is paid, while in the USA and Western Europe borrowers can requesttheir own credit history for free once a year.The loan with the interest rate range takes into account the following items:1. If the client has no credit history, then he/she will not be able to apply for aminimum rate. If the client requests the loan for the first time, there is a highprobability of an average or even maximum interest rate.2. Early repayment affects the cost of borrowing. This point has to be providedin the text of the credit agreement.3. From the moment of receiving the loan the borrower begins to “work" for thecredit history.

Even very responsible, law-abiding citizens can have arrearsbecause of bad organization. To avoid arrears you should make payments somedays before the date of payment. In case of force-majeure it is possible to leaveon the account the sum equal to one payment on the loan. At the same timeearly repayments are not welcome by banks.The main criterion of reliability of the borrower is a providing loan payments in time.92.2Neural Networks in Credit ScoringProbably the most complicated algorithm, and therefore less common for solv-ing credit scoring task is a neural network construction.Over the years, linear modeling has been the main modeling method in most areas, since optimization procedures are well developed for it.

In tasks where the linearapproximation is unsatisfactory (there are many such ones), the linear models workpoorly whereas neural network can reconstruct non-linear dependencies. Besides,neural networks cope with "curse of dimensionality" [19], which does not allow oneto model the linear dependencies in case of a large number of variables.In order to train neural network data scientist selects representative data and thenstarts training algorithm which automatically perceives data structure.

At the sametime the data scientist needs some set of heuristic knowledge about how to select andprepare data, choose the appropriate network architecture.Neural networks are attractive from an intuitive point of view as they are basedon a primitive biological model of nervous systems. In the future one believes thedevelopment of neurobiological models can lead to the creation of truly artificialintelligence. Meanwhile, already "simple" neural networks are a powerful weapon inthe arsenal of a specialist in applied statistics.Neural networks arose from researches in the field of artificial intelligence,namely, from attempts to reproduce ability of biological nervous systems to studyand correct errors, modeling low-level structure of a brain (Patterson, 1996).

Themain area of researches on artificial intelligence were expert systems in the 60th - the80th years. Such systems were based on high-level modeling of process of thinking(in particular, on representation that process of our thinking is constructed on manipulations with symbols). It soon became clear that such systems, although they couldbenefit in some areas, did not capture some of the key aspects of human intelligence.According to one view, the reason of this was that they were not able to reproducethe structure of the brain. To create artificial intelligence, you need to build a systemwith a similar architecture.The brain consists of a very large number of neurons connected by numerous10connections (on an average several thousand links per neuron, however this numbermay vary greatly).