Abstract - The article presents the way of applying cluster models to customer classification and managerial

decision on retaining the available clients and acquiring new ones. The objective of the research is to find out

the relevant techniques for building scoring models in different fields. The main research was testing the

hypothesis: if the number of point models is approximated in different spheres of activity, then the proposed

methods will be universal. To check this hypothesis the vector method of k-nearest neighbors support was

applied for decision making in the digital economy based on scoring models. In order to realize the principle of

customer classification and revealing the client categories with risk of quitting, the client’s classification model

was created. Moreover, a risk issue was shown in the example of fraud dynamic. Different fraud categories

were studied to define their features. On the basis of the model building results, the authors proposed some

recommendations on decision making in risk situations. The model shows how to retain existing clients and

Development and widespread use of computer

technologies have caused significant changes in

economies. Many activities increasingly rely on the

Internet. In many spheres (for example, insurance,

health, science, agriculture), information is

digitized, [1]. All they allow to accumulate and

create large volumes of data (Big Data), processing

the economy and society. The electronic toolkit

leads to radical changes in people's lives and is one

of the priorities for leading countries in economy,

including the United States, Germany, China, Japan

etc. The creation of new technologies including the

considering loan applications, the bank carries out a

thorough check up of potential customers for

possible financial risks. For each client, the

entrepreneur estimates the prospect of retaining the

person as a client for a maximum possible time. For

this purpose, the risk of possible client loss is again

assessed. Prior to launching a product to the

market, an entrepreneur estimates the possible

buyers and assesses all the risks.

into groups. It assumes that people with similar

social indicators behave the same way. Today

scoring is used in banking, marketing, insurance

and many other spheres of business activity.

Literally, "scoring" means counting of scores. This

article shows what kind of points modern analysts

consider and what for they use them.

transmission and submission of data electronically,

[4]. Richard Heeks in his research on Information

and Communication Technologies for

Development, points out that these technologies

create new opportunities in the digital sphere: an

entrepreneur or a company can optionally use the

digital system in their activities, [3]. This process

may include datafication (implementation of

storage technologies for large arrays), digitization

(conversion of all parts of the value chain from the

exclusively to the events that are currently

underway and the unfinished transformation of all

sectors of the economy due to the digitization of

information by using computer technology. It is

explained by the fact, that the informational

technologies enable payments on line, concluding

contracts have become known along with

blockchain and cryptocurrency technology. Now

they are part of reality. But smart contracts are

much wider than algorithms that translate tokens.

They will change the legislative reality, lawyers’

work and business life, [8]. Access to data in digital

economy became a key factor in product

development and innovations, and data collection

and their use by the third parties causes the

controversial issues concerning protection of the

regulation therefore state policy can consider not

fully the growing competition in various industries.

regulation of digital intellectual economy" K.M.

Belikov says, "to guarantee the open markets,

innovations, quality and efficiency and also

freedom of choice for consumers, the effective

competition has to be protected from restrictions",

[9].

the conditions of digital economy embrace the

sphere of digital economy come down to the same

conclusion, that it is necessary to provide such a

legal regime which would enable fee development

of innovations and prevent possible risks. Since

there exists a risk of disability to precisely predict

the development of innovations in digital economy,

the new legislation must be flexible and take into

account a great number of data, [11]. Around the

world the global financial institutions face various

Bolton's dissertation "Logistic regressions and their

application in credit scoring", (2009) the concept of

credit scoring for banking in South Africa is

considered. It points out the methods of

constructing a scoring model with emphasis on the

logistic regression method, [17]. The thesis of

Matthias Kremlin "Adaptive models and their

application in credit scoring", (2011) analyzes

methods of constructing predictive models in the

conditions of drift and data retention. A new

 to discuss different ways of determining risk

past similar contacts. It should be noted that

scoring is not an ideal financial risk analysis

program but it helps to quickly and accurately

make managerial decisions when working with big

data, [22].

According the scoring tasks, all scoring models are

divided into three categories, demonstrated in

Figure 1.

individual. If, after entering all the answers in the

program, the loan officer/mortgage banker claims

that the scoring has been completed, this means the

completion of the bulk of the analytical

verification. Next, the person’s loan application

goes to the security service, where bank experts

complicated and considers many factors when

information. Each of his/her characteristics has

points of certain value. After checking the

reliability of these data and summing up the scores,

a decision is made on the solvency of the potential

borrower and, based on this value, on the issuance

or non-granting of a loan. The value of the

“passing” score depends on the loan product.

Fig. 1: Scoring model classification according to

the defined tasks

Source: compiled by the authors

circumstance is found without obvious criminal

signs, but it requires explanation. For example, two

loan applications have the same address of

residence and home phone number. Perhaps these

are people who live in a civil marriage, or maybe

it's a dummy phone and address. In this case, the

verifier calls one of the clients and finds out

whether these people know each other, clarifies

some parameters in order to understand the

credibility of the client’s words. Risk management

even know about. However, it cannot foresee all

situations.

on the mathematical methods for building these

models, [28]. Among the statistical methods are

popular discriminant analysis, linear regression,

logistic regression and decision tree. Other methods

originate from mathematics: mathematical

programming, neural networks, genetic algorithms

Source: Compiled by the authors

discriminant analysis is a method for classifying

objects into predefined categories. Its main idea is

to find a linear combination of explanatory

variables that would best categorize objects. By

separation, it is best understood as one that ensures

the maximum distance between the average of

these categories. The score is calculated as a linear

which buyers evaluate a product in this category.

Use quantitative marketing research methods (such

as surveys) to collect data from a group of potential

buyers judging by their preferences for the

characteristics of the goods. Data for various

products are encoded and entered into a statistical

system, such as R, SPSS, or SAS.

discriminant function and determine the statistical

when the dependent variable has only two states.

The multiple discriminant method is used when the

dependent quantity has three or more qualitative

states. SPSS uses Wilk's Lambda or F-stat in SAS

to test significance. The most common method of

evaluating solvency is to divide the available data

into estimates and verification or deferred data.

Evaluation data is used to construct the

discriminant function. The deferred data is used to

construct a classification matrix, which indicates

the number of correctly and incorrectly classified

degree of difference between them. Dimensions

must be determined by the researcher. This requires

subjective judgment and is often a daunting task.

the simplest scoring method. In the case of two

categories, it is equivalent to the linear discriminant

analysis method and expresses the dependence of

one variable (dependent) on the other

independent variables;

coefficients that are found by the least squares

method;

accuracy of the forecast, it is impossible to

understand the reasons why a decision was made.

It’s the main disadvantage of the use of neural

networks for scoring models development.

method for classifying objects is based on a metric

that determines the similarity between the data.

Initially, training data is divided into classes, then

the evaluated data are entered and the similarity

The smaller the class, the less k is chosen.

Moreover, the result for the large k will not

necessarily be better. One of the advantages of this

method is the easy possibility to add new data

without changing the model, [32]. The

nonparametric nature of this method allows to work

with irrationalities in risk functions in the attribute

space. The absence of a formal method for

choosing k and the impossibility of a probabilistic

interpretation of the result, are the main

disadvantages of the method. These difficulties can

curve. Eight data sets have been studied (Table 1).

Support Vector Method and Logistic Regression

were the best in the studied eight data sets, [33].

There is no optimal scoring model for any

situation. The choice of model depends on the data

and the purpose of creating the model. In addition,

the best rating method will not necessarily be the

best in this situation.

the research, in the calculations, [34], it was applied

the set of scoring models as well as a probability

statistical model. The first model using the data of

real gym company help to build the system of

decision making for risk clients. And the second

one describing the fraud processes was based on

the biggest probability to leave this gym next

month. So, there are two risk groups at this gym.

The riskiest group is represented by the Cluster 1

and the next risk group is shown in the Cluster 5.

Fig. 3: Customer Classification Scoring Model

Source: Compiled by the authors

grandchildren having no high school diploma, [35],

[36].

Fig. 4: Risk clusters client’s characteristics

Source: Compiled by the authors