KR20090006489A - Toolkit of constructing credit risk model, method of managing credit risk using credit risk model construction and recording medium thereof - Google Patents

Abstract

A tool kit for constructing a credit risk model, a method for managing credit risk by constructing the credit risk model, and a recording medium are provided to quickly process and keep mass credit management data in spite of a small scale system, and offer various usability of a credit management system to users. A raw data database(110) stores accounting data of an enterprise or a person. A data manager(120) removes a missing value from the data selected from the raw data database and manages the data including no missing data. An analysis database storing unit(130) stores the data of the data manager. A model constructor(140) selects a variable of a credit evaluation mode from the data of the analysis database storing unit and constructs the credit evaluation model from the selected variable by using recursive logistic analysis. An analysis result storing database(150) stores an analysis result of the constructed credit evaluation mode. A model evaluator evaluates a credit level of the enterprise or the person.

Description

Toolkit of constructing credit risk model, Method of managing credit risk using credit risk model construction and Recording medium

The present invention relates to credit risk management, and more particularly, to establish a model for credit risk management of an enterprise or individual, and based on the established model, credit risk that can be applied to the distribution field as well as credit risk management work. Model kit, model of credit risk management through credit risk modeling and recording media.

In line with the development of the financial industry, the growing awareness of risk management has led to the need for regulatory capital by banks. Financial institutions have recognized the importance of risk management through several financial crises and are making great progress in their applied techniques.

In addition, in the industrial technology sector, competition among companies in the market has become fierce, making it difficult to predict a company's future creditworthiness. For financial institutions, how efficiently they can manage and transfer these credit risks is an important factor in their competitiveness, and the need for evaluation of credit risks is emerging as a method.

Credit risk means the likelihood that a borrower of a fund will not be able to repay a promised amount. A credit risk is a spread of baseline indicators such as credit rating and risk-free interest rate. This is reflected in the market price of spread and debt.

In the management of such credit risk, evaluation of various companies is carried out through a small number of specialized corporate evaluation agencies in Korea, and in major financial institutions, the establishment of a system for managing credit risk is approached in terms of time and cost. This is a difficult condition.

However, the credit risk management models established at each financial institution are implemented through outsourcing through professional workers rather than the financial institutions' own manpower. Professional personnel should combine not only professional financial engineering knowledge but also statistical knowledge, and perform their own development using programs for commercial statistical analysis and commercial development software, rather than specialized systems for model building.

Therefore, in order to develop a system for specialized credit risk management, the time required for considerable time and expert utilization is inevitably increased.

Even after the system is built, the system itself is often quite large. Therefore, in order to operate a separate credit management system, additional hardware construction and operation system construction must be involved, and even after the system is established, active maintenance and utilization of the system requires significant costs for additional professional employment and system operation. .

Therefore, the conventional system for managing credit risk is modeled only by a few experts with knowledge of credit risk, and the system is so large that it requires huge capital to operate it. There is a problem in that an additional expert is required for the use, and a considerable cost is required for this.

Therefore, the first problem to be solved by the present invention is to build a credit risk model that can quickly process and maintain a large amount of credit management data while giving a user a variety of credit management system while the system is small in size. To provide a toolkit.

In addition, a second problem to be solved by the present invention is to provide a credit risk management method by establishing a credit risk model that can perform credit risk management using the credit risk model building toolkit.

In addition, a third problem to be solved by the present invention is to provide a computer-readable recording medium for executing the credit risk management method through the credit risk construction on a computer.

The present invention to solve the first problem,

A raw material database for storing data on a financial statement of a company or an individual for credit evaluation, a data management unit for selecting data from the raw material database to remove missing values of the selected data, and managing data from which the missing values are removed; Analytical database storage for storing data of the data management unit, a model for selecting a variable of the credit rating model from the data of the analysis database storage, and build a model of credit rating model from the data by logistic regression analysis from the selected variable And a model evaluation unit configured to store an analysis result of the credit evaluation model constructed from the model construction unit, and a model evaluation unit evaluating the credit rating of the company or the individual by using the analysis result storage database unit. Toolkit for building risk models.

And, in order to solve the second problem, the present invention,

Selecting data to be analyzed from an original data database in which data calculated for corporate or individual credit rating is stored; independent variables which are variables affecting the credit rating in the selected data and credit rating according to the independent variables Designating a response variable that is a variable representing a result, calculating basic statistical information of the selected data using the independent variable and the response variable, and refining the data by detecting missing values in the independent variable among the selected data Selecting independent variables to perform the credit evaluation by performing basic statistics, correlations between the independent variables, and regression analysis based on the purified data, and using the selected independent variables, a prediction table for credit evaluation. Construct and recall according to logistic regression Building a credit rating model using the constructed prediction table; and estimating the credit rating by calculating the credit rating value using the constructed credit rating model. to provide.

In addition, the present invention to solve the third problem,

The present invention provides a computer-readable recording medium for executing a credit risk management method through a credit risk construction on a computer.

According to the present invention, there is no need to modify the data currently held in a special form for analysis of credit risk management, and various types of document data can be used according to the user's situation, and for small scale credit risk management The toolkit can handle large amounts of data quickly, easily handle complex tasks such as outliers and missing values in a graphical user interface environment, and provide various models for credit risk management according to variable selection analysis methods. It is effective to provide compatibility according to the bankruptcy prediction model that is widely used and to provide effective classification rules for corporate bankruptcy.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings.

However, embodiments of the present invention illustrated below may be modified in various other forms, and the scope of the present invention is not limited to the embodiments described below. Embodiments of the invention are provided to more fully illustrate the invention to those skilled in the art.

1 is a block diagram illustrating a toolkit for constructing a credit risk model according to the present invention.

Referring to Figure 1, the credit risk model building tool kit 100 according to the present invention is a raw material database 110, data management unit 120, analysis database storage unit 130, model construction unit 140 and The analysis result storage database unit 150 is made.

First, the raw data database 110 stores data on financial statements of a company or an individual for credit rating.

In the present invention, since the analysis result is stored in the user's work unit, there is no need to perform a separate data integration work to perform this. That is, various data sources of different formats can be loaded in one unit of work, and each data source loaded in one task can be stored under a separate data name.

On the other hand, in the present invention, since various types of data can be used, it is possible to perform an analysis on creditability by importing all easily used Excel files or various text type files of Microsoft Corporation. Can be.

The data manager 120 selects data from the raw data database 110, removes missing values of the selected data, and manages data from which missing values are removed.

To this end, the data manager 120 includes a data definition module 121, a data purification module 122, a data replacement module 123, a basic statistical analysis module 124, a correlation analysis module 125, and a regression analysis module 126. It can be made of).

The data definition module 121 designates an independent variable, which is a variable that affects the credit rating, and a response variable, which is a variable representing a result of another credit rating, in the data selected from the raw material database 110. By using the independent variable and the response variable, basic statistical information of the selected data may be calculated.

 Herein, the independent variable may refer to a risk factor related to the credit rating, and in addition to the independent variable and the response variable, the time series variable of the data, the size of the data, or the data Categories can be customized.

In addition, the basic statistics information calculated from the data definition module 121 includes the type of data selected from the source database 110, the role of each variable, the size of the data sample, the number of missing values, the average, the variance, and the five-digit data summary. May contain a value.

In detail, when using data to construct a model for corporate default, a variable representing the default can be defined as a response, and general financial statements can be defined as a credit factor. Can be. If you have categorical data, such as company size, that is divided into large or small businesses, you can specify it as a category, and if it contains variables that represent time, such as year, quarter, month, or day, you can specify it as a time variable. have.

The role of each variable designates the role of each variable of the data, and may designate a time, a size, or a category in addition to the aforementioned risk factor and response. .

In addition, the size of the data sample may be expressed in Arabic numerals, and information on the number of missing values (NMiss) is displayed.

Among the data prepared for the credit analysis, there may be data with extremely large or extremely small values. If such data is called a missing value and the missing values are used as is in performing a credit analysis, an error may occur in the reliability of the data. Therefore, before performing the actual credit analysis, the data identified as missing values in the selected variable are automatically searched and the total number of missing values NMiss is output.

The basic statistics information may include information about mean, variance, and standard deviation, and provides a standard error value that provides a reference for the accuracy of the mean of the data sample. It may include.

In addition, the basic statistics information may include data five-digit summary value, the data five-digit summary value is the minimum value (Minimum), the first quartile value (1st Quartile), the median value (Median), the third quartile value ( 3rd Quartile) and Maximum.

Here, the first quartile value (1st Quartile) means the range of the value corresponding to the lower 25% when the data is arranged in size order, and the median value is the middle when the data is arranged in size order. If N, the size of data, is odd, the (N + 1/2) th value is the middle value, and if N is even, the N / 2 th value and (N / 2) +1 th value The mean is the median.

And, the third quartile value (3rd Quartile) may mean the value of the top 25% when the data are sorted in size order.

The data purification module 122 detects the selected data within the independent variable and refines the data.

Data prepared for analysis may contain data that are extremely large or extremely small compared to other values. If these missing values are used as they are in the data analysis, they can produce incorrect results, so perform data cleansing that can be processed before entering the actual analysis. The data purification module 122 may immediately check the process of changing the attribute of the variable before and after the missing value is removed, and the result may be automatically stored.

For example, in the assessment of poor companies, the refining of data is to remove missing values to stabilize the distribution of the data and to increase the discernment of credit risk through the refined data. In addition, the user may perform a function on all or a variable selected by the user with respect to the data, and provide the user with a result of processing the modified data.

The data replacement module 123 replaces missing values detected by the data purification module 122 with replacement data.

That is, if the data to be used in the credit rating model has a missing value, the data replacement module 123 substitutes data preset by the user and adds data to the entire analysis data.

On the other hand, in constructing a credit rating model, if the ratio of the missing values is too large in the data to be analyzed, this means loss of information. In particular, if the occurrence of the missing value is at the structural level, the missing value itself contains sufficient information and should be used in the credit rating model.

In order to use the missing value in the credit rating model, the missing value should be replaced with another value. In the present invention, the average value of the observed values is calculated by using the mean replacement method, and the missing value is replaced with the calculated mean value, or the median is used. Alternatively, the missing values may be replaced with the calculated median by calculating the median of the observed values. The replaced missing value may be stored in the data purification module 122 in the form of new data.

In addition, the basic statistical analysis module 124 calculates basic statistics for each of the independent variables required to build the credit risk management model. The basic statistics may be calculated by the user when each variable is defined in the data definition module 121, and the basic statistics may be calculated after the data is purified in the data purification module 122.

Correlation analysis module 125 calculates the correlation coefficient between each independent variable required to build a credit rating model. In order to construct the credit rating model, the correlation coefficient between each independent variable and financial statements used as a risk factor is calculated, and the statistical significance of the calculated correlation coefficient can be verified and the verified significance result can be immediately retrieved. Can be. Through this, the relationship between various financial statements is verified and actual analysis can be made before establishing the actual credit rating model.

The regression module 126 compares the degree of influence of the risk factor, which is an independent variable, on the response variable, to see how the response variable is explained or predicted by other independent variables. The significance test for the influence on the dependent variable is performed. This can be applied to market model, capital market line model and capital asset pricing model such as stock market line model in the real financial field.

The analysis database storage unit 130 stores data that has undergone statistical significance verification by the regression module 126.

The data stored in the analysis database storage unit 130 selects a variable of the credit rating model in the model construction unit 140, and builds a credit rating model from the data by logistic regression or inverse regression analysis from the selected variable. .

Meanwhile, the model building unit 140 may include a segmentation module 141, an independent variable selection module 142, a logistic model module 143, and a calibration module 144.

The segmentation module 141 provides a classification rule using a unique discrimination model. This constructs a suitable model for classifying individual companies according to a predefined group using various financial statements of a company or an individual, and provides a classification rule based on the constructed model. The classification rules established allow you to determine groups for an individual company, and also make classification predictions for new individuals or companies that have not yet been grouped. Or, it can be classified as bad, financial institutions use the data of such individuals or companies to determine whether there is a good or bad according to each credit status, insurance companies, etc. segmentation (segmentation) to prevent the departure of customers ) It is used for building and assigning a customer grade.

Independent variable selection module 142 can be used to build a credit rating model using a very large number of factors in establishing a credit risk management model for a company or an individual. Using all of them in one model can reduce the efficiency of the analysis and can make the built model less useful.

The present invention provides criteria for selecting appropriate risk factors in order to increase the efficiency of credit rating model construction and the utilization of the model. By presenting statistics for selecting various risk factors based on the generalized selection model frequently used to build a model of credit rating, the user can select the appropriate risk factors, that is, independent variables, by checking the basic statistics provided by the present invention. have.

Logistic model module 143 can identify the major risk factors affecting credit risk for predicting the creditworthiness of an individual or a company in constructing a credit rating model of an individual or a company, and the statistical significance and stability of the constructed model This can be confirmed through the basic statistics. And, using the independent variable according to the risk factors to extract the cut off point (cut off point) for the credit rating to predict the credit rating.

If the calibration module 144 estimates the creditworthiness using the logistic model module, the calibration module 144 evaluates the creditworthiness based on the data on the financial statements of the company or the individual, and thus may show a certain difference from the actual company or individual creditworthiness. In order to re-adjust these differences to the actual creditworthiness of the company or individual, the reestimation of the creditworthiness is performed using the calibration module.

In order to perform reestimation of the creditworthiness, a regression method using a nonparametric regression model may be used, and in particular, a creditworthiness re-estimation may be performed by using a locally weighted robust bust regression method (LOESS). Through this reestimation process, the credit rating value close to the actual credit rating is estimated.

The credit rating value is stored in the analysis result storage database unit 150 so that the user can utilize the credit rating value whenever necessary.

In addition, the data stored in the raw material database 110, the analysis database storage unit 130 and the analysis result storage database unit 150 can be stored in the form of XML document, so that the user can easily Can be used.

Figure 2a shows the data selected from the source database applied to the present invention.

Referring to FIG. 2A, first, as described above, data stored in an original data database may be retrieved by a user's selection.

In the screen of the data sheet loaded by the user's selection, the first row at the top means each variable of data, and each column shows actual data according to each variable.

2B illustrates variable information of data in the selected data.

Referring to FIG. 2B, a region corresponding to 'Role' denotes a role of a variable, and a role of a variable may be designated for each variable. As described above, the independent variable and the dependent variable, which is a response variable representing a result value corresponding to the independent variable, may be designated.

In FIG. 2B, the independent variable means 'Risk Factor', and the response variable, which is the dependent variable, is designated as 'Response'. In addition, time is 'Time', data size is 'Size', data category 'Category', etc. Can be specified by the user, and by specifying the variable information of the data, it is possible to perform the actual credit rating.

Figure 2c shows the basic statistics information output by the selected variable.

The type of data selected from the data selected according to the variables specified in FIG. 2B, the role of each variable, the size of the data sample, the number of missing values, NMiss, mean, and variance. (Variance) Standard and data five-digit summary values can be included.

FIG. 3A illustrates a data purification screen applied to the present invention, and FIG. 3B illustrates a data purification result of data selected by FIG. 3A.

Referring first to Figure 3a, it is determined whether to refine the data for each independent variable in the selected data sheet. In this case, select 'Cleaning' as the independent variable to clean up data and 'None' as the independent variable for data not to be purified.

Then, as shown in FIG. 3B, the data purification result can be confirmed according to the selected result.

Referring to FIG. 3B, although there was no data corresponding to missing values before data cleaning (Uncleaned), 13 missing values increased after data cleaning (cleaned).

In addition, the box plot shows the distribution of the data after data refinement.

In the box diagram, line 311 represents the median of the selected data, line 312 represents the minimum of the data corresponding to the lower 25% of the median of the selected data, and line 313 represents the selected value. The maximum value of data corresponding to the top 25% from the median value of the data is displayed.

Here, the value of the line 312 which is the minimum value of the data corresponding to the lower 25% value from the median value of the selected data and the line 313 which is the maximum value of the data corresponding to the value of the upper 25% from the median value of the selected data ) Range is referred to as Inter Quartile Range (IQR) 310.

  In addition, a line corresponding to the value of IQR × 1.5 from the line 312, which is the minimum value of the data corresponding to the lower 25% of the median value of the selected data, is referred to as the minimum threshold line 320. The line corresponding to the value of IQR × 1.5 from the line 312, which is the maximum value of the data corresponding to the upper 25% value of the median of the selected data, is referred to as the maximum limit value line 330.

Then, a value out of the range of the minimum limit line 320 and the maximum limit line 330 is called a missing value, and the missing value has such a value because the missing value has an extremely large value or an extremely small value compared to other values. If used as is, it can lead to incorrect results.

Therefore, when 'missing' is detected after selecting 'Cleaning' as an independent variable to refine such missing value data, more reliable basic statistics information can be obtained by confirming the basic statistics information excluding the missing values as shown in FIG. 3B.

FIG. 4A illustrates replacing the missing value of FIG. 3B.

Referring to FIG. 4A, when a missing value exists in FIG. 3B, in order to use the missing value with respect to the variable in which the missing value exists, the missing value should be replaced with another value. The missing values may be replaced by a mean imputation by calculating an average of the estimated values, or by substituting a median value observed by a median imputation method. The replaced missing value is to be stored as an average value or a median value in the form of new data.

FIG. 4B shows information of data in which missing values are replaced by FIG. 4A.

Referring to FIG. 4B, when the average value of the values observed as the mean imputation in FIG. 4A is calculated to replace the missing value corresponding to the independent variable of X4, all of the 13 missing value values are replaced by the average value, so that the missing value exists. You can see that it does not. In addition, the value of the variance, which represents the degree of variance of the data, decreased from 146.94 to 127.45, indicating that the data were less deviated.

Figure 5 shows a correlation analysis of the data applied to the present invention.

The present invention can perform a process of searching for an appropriate variable for credit risk management through correlation analysis of data.

First, by selecting a plurality of independent variables for correlation analysis of data by the user and performing correlation analysis between the independent variables, independent variables suitable for credit risk management can be selected.

In addition, the basic statistical information 510 of the purified data together with the correlation analysis table may be confirmed, and the statistical significance 520 of the correlation coefficient may be simultaneously verified.

Referring to FIG. 5, five independent variables are used for correlation analysis. First, basic statistical information of data corresponding to five independent variables may be checked, and a correlation analysis diagram represented by a matrix may be obtained. You can check it.

The correlation analysis diagram displayed as a matrix may define a user-defined significance level. In the present embodiment, the significance level is defined as 0.05, and when the significance level is less than 0.05, the p-value is shown in white.

It should be noted that in the results of the correlation analysis, the coefficient among the correlation coefficients between variables is the largest value. Referring to FIG. 5, it is a correlation coefficient between the total capital recurrence rate represented by the first variable X1 510 and the equity return ordinary income rate represented by the second variable X3 530. Correlation coefficients between two variables are closely related, and if the two variables are used as independent variables in the logistic regression model for the default prediction model, the problem of multicollinearity may occur.

In general, when the correlation coefficient between the independent variables is 0.7 or more, it can be suspected multicollinearity, and when the correlation coefficient is 0.9 or more, it can be seriously suspected. If multiple collinearity occurs, normal regression analysis cannot be performed and errors can be derived that result in incorrect inferences from inferences about individual regression coefficients.

As shown in FIG. 5, since the correlation coefficient between X1 510 and X3 530 exceeds 0.9, two variables X1 (510) and X3 (530) are needed to perform logistic regression analysis for actual model construction. ), Only one variable should be used as the risk factor.

Figure 6 shows the analysis results for the selection of independent variables by logistic regression analysis applied to the present invention.

Referring to FIG. 6, it can be seen that the independent variables, which are candidate variables to be used for the credit rating, are selected one by one and output the results. This is calculated by the logistic regression module.

According to the logistic regression module, the statistics of the independent variables used as the selection criteria are the Akaike Information Criterion (AIC), the Schwarz Bayesian Criterion (SBC), and the Accuracy Ratio. Among them, AIC and SBC can be independent variables suitable for credit analysis with smaller values, and higher accuracy ratio is suitable for credit analysis.

Observation 610 in Figure 6 refers to the total number of data for each independent variable for evaluating the bankruptcy of the company used in the regression module.

Missing (620) means the number of data corresponding to missing values for each independent variable used in the regression module.

The default count (630) refers to the number of data represented as actual defaults in default variables, which are response variables in the regression model for predicting default. In the example data used in the analysis, out of a total of 99 companies, two are bankrupt.

을 의미한다. 여기서,

는 하기 수학식 1에서 절편에 해당하는 값으로 입력될 독립변수(Risk Factor)가 없을 경우에 적용되는 값을 의미한다.Estimate (640) is an estimate of the regression coefficient used in Equation 1 below in the regression analysis of each independent factor when p default probability is p.

Means. here,

Means a value applied when there is no independent factor to be input as a value corresponding to the intercept in Equation 1 below.

In addition, the above-mentioned Akaike Information Criterion (AIC) (650) is a measure used to compare the credit model fit between the fitted models, the smaller this value, the better the fit of the credit model, This is the preferred model. Here, the AIC may be calculated as in Equation 2 below.

는

번째의 관찰치의 빈도를 의미하고,

는

번째 관찰치의 추정된 부도확률,

는 반응변수의 범주-1 으로 여기서는 1을 의미할 수 있고,

는 독립변수의 수를 의미하고, 상기의 단순 로지스틱 회귀모형에서는 1을 나타낸다.In equation (2)

Is

Means the frequency of the first observation,

Is

Estimated default probability of the first observation,

Is category-1 of the response variable, which can mean 1,

Denotes the number of independent variables, and 1 in the above simple logistic regression model.

In addition, the Schwarz Bayesian Criterion (SBC) 660 is a measure used to compare model fits between fitted models, similar to the AIC, and the smaller the value, the better the fit of the model. It is a preferred model. Meanwhile, the SBC value which is the Schwarz Bayesian criterion may be calculated as in Equation 3 below.

는

번째의 관찰치의 빈도를 의미하고,

는

번째 관찰치의 추정된 부도확률,

는 반응변수의 범주-1 으로 여기서는 1을 의미할 수 있고,

는 독립변수의 수를 의미하고, 상기의 단순 로지스틱 회귀모형에서는 1을 나타낸다.Similarly, in equation (3)

Is

Means the frequency of the first observation,

Is

Estimated default probability of the first observation,

Is category-1 of the response variable, which can mean 1,

Denotes the number of independent variables, and 1 in the above simple logistic regression model.

In addition, the accuracy ratio (Accuracy Ratio) 670 is a measure between the value of the response variable (here, default) and the probability of default predicted by the model. The larger the value, the better the fit of the model.

The measuring method of the accuracy ratio is shown in Equation 4 below.

와

라 하고 하여 상기의 수학식 4와 같이 정확도 비율이 측정될 수 있다.In the method of measuring the accuracy ratio based on Equation 4, two data are extracted from all the data, and in this case, the two values should have different response variables. In other words, if the value of one variable is a default, the value of the other variable must not be a default. Consider these two data as a pair. If we define t as possible for all possible pairs, we identify the default probability calculated by the model for each pair. In this case, if the predicted probability of default of the non-default company and the default company in each pair is in the same direction, that is, if the probability of default of the defaulted company is greater than that of the non-default company, it is called a matched pair. If not, that is, if the default probability of a company that is defaulted is less than the default probability of a company that is not defaulted, it is called a mismatch pair. The total number of matched and unmatched pairs, respectively

Wow

The accuracy ratio can be measured as shown in Equation 4 above.

7A illustrates independent and response variables when constructing a credit rating model using the logistic regression model of the present invention.

In practice, the ultimate goal of the credit valuation model is creditworthiness, for example, corporate defaults.

When considering a company's financial statements for a particular company, the statistical model is to determine whether the company is in default (y = 0) or not (y = 1). Logistic regression analysis can be said to be a suitable method as a statistical model for data with two values of response variable y.

The logistic regression model can be used to identify independent variables that have a significant impact on the bankruptcy of a particular company, and the degree of influence on the bankruptcy can be compared through relative comparisons.

In addition, it is possible to estimate the predicted default probability of each company, which is the final purpose of the model, and to provide a table for performing the evaluation of a suitable model and finding an appropriate cutoff point for the default.

For example, we select three variables from a database of individuals or companies, and use the model of logistic regression as the default predictive model.

This means that if the model is fitted according to the characteristics of the data, all the results can be confirmed on the result screen. The suitability of the logistic regression model constructed from each result, various basic statistics related to the model fit, a table to confirm the model execution, a table to search the cutoff points to determine the creditworthiness, and finally all the companies used in the analysis. Estimated probability of company-specific default forecasts for the data can be identified.

First, referring to FIG. 7A, a user designation may be performed for analysis in a logistic regression analysis window.

First, the user can specify multiple data sets that can be executed in one project. If a plurality of datasets are called in the process of selecting data, the names of the called datasets are output. Among them, the user can select a data set for building a model (Select Data Set) 711. If there is only one data set called in the project, no special designation is required.

Second, a plurality of models can be constructed in one project. In this case, the constructed model can be designated separately, and for this purpose, a name of the model is designated (Input Name) 712.

Third, a response variable of the logistic regression model to be constructed is designated (Target Variable) (713). For example, if you want to build a model for corporate bankruptcy forecasting, specify a variable that indicates whether or not the corporate bankruptcy exists. This can be done by specifying response variables when defining the data.

Fourthly, the independent variable of the logistic regression model to be constructed is designated (714). If you want to build a bankruptcy forecasting model, you can specify each financial statement as an independent variable, which can be specified in a graphical user interface environment.

Fifth, when outputting basic statistics and correlation analysis results for independent variables used in model construction, the basic statistical information may be output by selecting Descriptive Statistics 715.

Sixth, an intercept term 716 may be designated when an intercept is added to a logistic regression model to be constructed. If this is not specified, the logistic regression model without the intercept term is used to determine whether the model fits.

After specifying the six terms, the output analysis results can be divided into a total of five areas.

First, model information can be used to determine whether an error occurred or converged normally in the estimation process of the model. If the model estimation is made normally, 'Iteration finished' message can be checked without any special error message. If a problem occurs in the estimation process, an error message is automatically displayed.

This can confirm the model convergence results of the logistic regression model as shown in Figure 7b.

Second, with Parameter Estimate, it outputs values related to the estimates for each regression coefficient in the fitted logistic regression model. In this case, the significance of each independent variable can also be checked.

라 할 때 모형 구축을 위한 로지스틱 회귀모형은 하기의 수학식 5와 같다.In detail, in the logistic regression model for building the probability of default probability model, Y, the response variable, indicates whether there is bankruptcy by company. If the firm is bankrupt, y = 1; if normal, y = 0. And the probability of default of an individual company according to the value of a given independent variable

In this case, the logistic regression model for constructing a model is represented by Equation 5 below.

들에 대한 추정치는 뉴톤-랩슨(Newton-Raphson) 방법을 이용하여 하기의 수학식 6에서 연달아 연산되는

과

가 거의 일치할 때까지 반복 수행하여 수렴하는 값이 각 회귀계수에 대한 추정치가 된다.Regression coefficient from equation (5)

The estimates for these fields are calculated in succession using Equation 6 below using the Newton-Raphson method.

and

Repeatedly converge until the values are nearly identical, and the convergence value is an estimate for each regression coefficient.

Third, various statistics are provided to confirm the goodness of fit of the raw data which is the logistic regression model fitted with the Model Fit Information.

As can be seen in Figure 7c it can confirm the risk factors of the logistic regression model of the present invention, that is, statistical information about the independent variable. In other words, in the case of applying to a default prediction using a model that is fitted using the provided statistics, it is possible to confirm the degree of suitability of the independent variable for the model.

This information includes information such as -2 loglikelihood, AIC, SBC, chi-square, p-value, and accuracy ratio.

Among them, AIC and SBC are used to compare the relative goodness-of-fit with models constructed using different independent variables. The smaller this value is, the better the model fits.

The chi-square and p-value are used to test the null hypothesis that all regression coefficients are zero in the fitted logistic regression model. The accuracy ratio is a measure of how well the fit results explain the actual bankruptcy and has a value between 1 and 0. The larger the value, the better the model fit.

에 검정결과 p-value가 0.01 보다 더 작으므로 귀무가설을 기각할 수 있다. 즉, 적합된 모형이 실제 자료를 적절히 설명하고 있다고 할 수 있다. 이는 도 7d를 살펴보면, 정확도 비율(Accuracy Ratio)을 살펴보았을 때 0.9355로 1에 가까우므로 모형 적합이 적절하다고 할 수 있다.This is a null hypothesis on the model suitability, looking at the model fit according to the logistic regression analysis shown in Figure 7d

As a result of the test, the p-value is less than 0.01, so the null hypothesis can be rejected. In other words, the fitted model adequately describes the actual data. Referring to FIG. 7D, the model fit is appropriate because it is close to 1 as 0.9355 when the Accuracy Ratio is examined.

Meanwhile, the -2 loglikelihood statistic may be calculated as Equation 7 below by multiplying the log likelihood function of the currently estimated model by -2.

는

번째의 관찰치의 빈도를 의미하고,

는

번째 관찰치의 추정된 부도확률을 나타낸다.In Equation 7

Is

Means the frequency of the first observation,

Is

It represents the estimated default probability of the first observation.

를 위한 검정통계량을 나타낸다. 로지스틱 회귀분석에서 개별 회귀계수들에 대한 검정은 왈드 검정(wald test)을 사용한다. 이 때 검정통계량은 이미 연산된 추정치와 추정치의 표본오차를 이용하여 하기의 수학식 8과 같이 연산될 수 있으며, 귀무가설하에서 자유도가 1인 카이제곱 분포를 따른다. 자유도가 1인 카이제곱 분포로부터 연산된 검정통계량을 이용하여 p-value를 연산할 수 있다.And chi-square is the test for the population regression coefficient with the estimated regression coefficient.

Assay statistics for. The test for individual regression coefficients in logistic regression uses the wald test. In this case, the test statistic may be calculated using Equation 8 below using the estimated value and the sample error of the estimated value, and follow the chi-square distribution having 1 degree of freedom under the null hypothesis. The p-value can be calculated using the test statistic calculated from the chi-square distribution with 1 degree of freedom.

Fourth, the Classification Table is a table that provides numerical values for preparing a positive classification graph in the form of a table. That is, the classification table and the positive classification graph may be used complementarily.

The Classification Table uses the results of a fitted default probability model to search for appropriate cutoff points for predicted default probability, and evaluates how accurate the fitted model produces in predicting actual default probability. A table that provides various numbers used to And the static classification graph is a graph showing the classification table at a glance.

The classification table according to the invention is as shown in 7e. Referring to FIG. 7E, a probability of a cutoff point for predicting whether an actual bankruptcy is predicted from the calculated default predictive probability is provided in 0.005 units. For example, when the cut-off point is 0.020 in FIG. 7E, if the default probability is 0.02 or more with a default probability calculated as a result of model fitting, the default value is determined as default.

The determination result can be summarized as shown in the classification table of Table 1 below.

According to the determination result classification table of Table 1, a case of predicting a company as an actual bankruptcy is one case and corresponds to an accurate event in the classification table.

In addition, as a result of the judgment, 87 companies predicted that they were normal, which corresponds to accurate non-events in the classification table.

On the other hand, one case of incorrectly predicting a company that is actually defaulted is a case of Incorrect Non-Event in the classification table.

In addition, six cases of incorrectly predicting firms that are actually normal as a result of the judgment correspond to an Incorrect Event in the classification table.

 Based on the classification table, the probability may be calculated as follows.

First is Percentage Correct, which is the probability of accurately predicting a bankruptcy from normal to normal.

Second, the sensitivity is the probability of predicting a real bankruptcy company as a bankruptcy.

The third is the specificity, which is the probability of predicting a normal company as normal.

Based on Table 1, the accurate prediction rate, sensitivity, and specificity are summarized in Table 2 below.

Cut-off point for increasing various frequencies calculated based on Table 1 and Table 2 and three probabilities calculated from the frequencies, Percentage Correct, Sensitivity, and Specificity, in increments of 0.005 Each of the three probabilities is provided through a correct classification plot. The normal classification graph is as shown in FIG. 7F. Referring to FIG. 7F, an optimal cutoff point may be obtained by using the classification table of FIG. 7E.

FIG. 7G illustrates a prediction table generated based on the correct classification graph of FIG. 7F.

Referring to FIG. 7G, a prediction table is a table that summarizes predicted default probability and actual default default value calculated from a model constructed for all individual companies of the data set.

This is a table showing the actual default and forecast default probability for 20 companies, which are part of the 99 company data used in the analysis. In FIG. 7G, an actual value of a company having an ID of 10 corresponds to 1, which indicates a bank default. In addition, since the predicted default probability (Predict) is 0.8841 from the model for this company, it can be seen that the very close default probability has been calculated.

On the other hand, in the case of a company with an identifier of 16, the company whose actual value is 0 is normal, but the predicted default probability is 0.114466, which is relatively larger than other predicted values.

This means that if the cutoff point is set to 0.02 as described in the classification table, 0.114466 is much higher than 0.02 and is incorrectly predicted as a bankrupt enterprise.

라 할 때, 상기 추정치를 이용하여 적합된 로지스틱 회귀모형은 상기 수학식 6과 같다.The calculation of the predicted probability for an individual company is performed using the estimated logistic regression coefficients. Estimates from Logistic Regression Models

In this case, the logistic regression model fitted by using the above estimate is as shown in Equation (6).

The predicted default probability p for each individual company from Equation 6 is calculated by Equation 9 below.

8A illustrates a data set designation screen for performing reverse regression analysis by the calibration module of the present invention.

As shown in FIG. 8A, a data set and a model name for performing reverse regression analysis are designated by the calibration module. On the other hand, when various models are created based on various conditions in the analysis, it is possible to designate each model to be calibrated.

In addition, the number of sections to be divided can be specified. This will re-estimate the overall predictive probability according to the divided interval, and output a more realistic graph. The minimum value of the divided section may be 20, and the maximum value may be 50.

8B shows an option designation screen for reverse regression analysis by the calibration module of the present invention.

Referring to FIG. 8B, it can be seen that this shows the options that can be specified for the nonparametric regression method. The non-linear regression method or the LOESS method described above can be selected.

The LOESS method has a variety of options to choose from: the order of the polynomial, the function used to determine the weight, the number of iterations required for the analysis, and the smoothing parameter. The value of the smoothing parameter can vary from 0.20 to 0.78. The number of repetitions required for the analysis can be specified from 1-5.

8C shows usage data of the regression analysis of the present invention.

Referring to FIG. 8C, if the interval is divided into 20 sections in the selected option when the analysis is performed, the value of the independent variable X1 will be inputted from 1 to 20, and the y value of the response variable is estimated by the logistic regression analysis. The probability is divided into 20 intervals to represent the actual default probability of that interval.

FIG. 8D is a graph outputting the calibration result according to FIG. 8C.

Referring to FIG. 8D, it can be seen that the prediction probability by the predictive model of bankruptcy and the fitting probability by LOESS are simultaneously shown.

As a result of applying the LOESS method, it can be seen that it is re-expressed in the form of a smoother function, that is, a monotonically increasing function as a whole, than the prediction probability by the existing prediction model.

9 is a flowchart illustrating a credit risk management method through establishing a credit risk model of the present invention.

Referring to FIG. 9, first, data to be analyzed is selected from an original data database in which data calculated for corporate or individual credit evaluation is stored (step 910).

That is, a user selects data to be directly evaluated from a database in which data calculated for credit rating of a company or an individual is stored. In the present invention, various types of data can be used. You can import all types of files and perform a credit analysis.

In the selected data, an independent variable that is a variable that affects the credit rating and a response variable that is a variable representing a result of the credit rating according to the independent variable are designated, and the import is imported using the independent variable and the response variable. The basic statistical information of the collected data is calculated (step 920).

It specifies the independent variables that affect the credit rating in the data selected from the source database and the response variables that are variables representing the results of the credit rating according to the independent variables.

Herein, the independent variable may mean a credit risk variable related to the credit rating, and the user may designate a time series variable of data, a size of data, or a category of data in designating the variable.

The basic statistical information may include the type of data selected from the source database, the role of each variable, the size of the data sample, the number of missing values, the average, the variance, and the five-digit summary value of the above-described data.

In operation 930, the missing value is detected within the independent variable among the selected data.

In the case of poor companies, if the selected data contain missing values, the data is replaced by data set by the user and added to the entire analysis data.

In the present invention, for this purpose, the values observed as missing values are replaced with the mean value of selected data by the mean replacement method, or the missing values are calculated by calculating the median value of the observed values using the median replacement method.

Based on the purified data, an independent variable for performing the credit evaluation is selected by performing basic statistics, correlation between the independent variables, and regression analysis (step 940).

Since the basic statistic is calculated based on the refined data, it is possible to know the distribution state of the selected data before and after the data refinement, and to analyze the correlation of each independent variable necessary to build a credit rating model. By calculating the correlation coefficient between the financial statements, which are independent variables used to construct the credit rating model, the statistical significance of the calculated correlation coefficient is verified and the verified significance results are retrieved.

Then, a regression analysis is performed to see how the response variables are explained by other independent variables, and the degree of influence of the risk factors as independent variables on the response variables is compared. Perform a significance test of influence. This may be applied to capital asset pricing models such as market models, capital market line models and securities market line models in the real financial sector.

Next, a prediction table for credit rating is constructed using the selected independent variable, and a credit rating model is constructed using the configured prediction table (step 950).

In building a credit rating model, a logistic regression analysis may be performed, or a regression analysis may be performed after the logistic regression analysis to build a credit rating model.

The credit rating model uses these independent variables to search for appropriate cutoff points for credit ratings, and evaluates the various values used to assess how accurate the fitted model produces in predicting actual creditworthiness. You can build a credit rating model by using the prediction tables provided.

Meanwhile, the credit rating model may include a logistic rating model.

In the logistic evaluation model, the user can specify a plurality of data sets that can be executed in one project. In the process of selecting data, a plurality of data sets can be called. Then, you can specify a model of creditworthiness and name the model based on the called dataset.

Then, the response and independent variables of the logistic regression model to be constructed are designated to model information on whether an error occurred or converged normally in the estimation process of the model, parameter estimation, and the original logistic regression model. The logistic evaluation model provides a classification table that provides the model suitability information that provides various statistics to confirm the goodness of fit, and the numbers for creating the static classification graph in the form of a table.

Finally, the credit rating value is calculated using the constructed credit rating model to estimate the credit level (step 960).

That is, it is possible to estimate whether the credit quality is poor according to the cutoff point calculated by the established credit rating model.

In addition, the present invention may re-estimate the credit rating value using the nonparametric regression model, and may evaluate the credit rating using the re-estimated credit rating value.

When estimating the creditworthiness using the logistic model, the credit rating is evaluated based on the data of the financial statements of the company or the individual, and thus may show a certain difference from the actual company or individual creditworthiness. In order to reconcile these differences with the actual creditworthiness of the company or individual, credit estimates are performed.

In order to perform the reestimation of the credit rating, a reverse regression method using a nonparametric regression model may be used, and in particular, a credit re-estimation may be performed by using a locally weighted robust regression method (LOESS). Through this reestimation process, the credit rating value close to the actual credit rating is estimated.

Such a credit rating value is stored in the analysis result storage database so that the user can utilize the credit rating value whenever necessary.

The present invention can be embodied as code that can be read by a computer (including all devices having an information processing function) in a computer-readable recording medium.

The computer-readable recording medium includes all kinds of recording devices in which data is stored which can be read by a computer system. Examples of computer-readable recording devices include ROM, RAM, CD-ROM, magnetic tape, floppy disks, optical data storage devices, and the like. The computer readable recording medium can also be distributed over network coupled computer devices so that the computer readable code is stored and executed in a distributed fashion.

Although the present invention has been described with reference to the embodiments shown in the drawings, this is merely exemplary, and it will be understood by those skilled in the art that various modifications and equivalent other embodiments are possible. Therefore, the true technical protection scope of the present invention will be defined by the technical details of the appended claims.

1 is a block diagram of a toolkit for credit risk model construction according to the present invention.

Figure 2a shows the data selected from the source database applied to the present invention.

2B illustrates variable information of data in the selected data.

Figure 2c shows the basic statistics information output by the selected variable.

Figure 3a shows a data purification screen applied to the present invention.

FIG. 3B shows the data purification results of the data selected by FIG. 3A.

FIG. 4A illustrates replacing the missing value of FIG. 3B.

FIG. 4B shows information of data in which missing values are replaced by FIG. 4A.

Figure 5 shows a correlation analysis of the data applied to the present invention.

Figure 6 shows the analysis results for the independent variable selection by the logistic regression analysis applied to the present invention applied to the present invention.

Figure 7a illustrates the independent and response variables of the credit rating model using the logistic regression model of the present invention.

Figure 7b shows the convergence results of the logistic regression model of the present invention.

7C shows statistical information on independent variables of the logistic regression model of the present invention.

7D shows the model fit of the logistic regression of the present invention.

7E shows a classification table applied to the present invention.

FIG. 7F illustrates a graph showing the classification including the cutoff point of FIG. 7E.

FIG. 7G illustrates a prediction table generated based on the classification graph of FIG. 7F.

8A illustrates a data set designation screen for performing reverse regression analysis by the calibration module of the present invention.

8B shows an option designation screen for reverse regression analysis by the calibration module of the present invention.

8C shows usage data of the regression analysis of the present invention.

8D is a graph of a calibration result output according to FIG. 8C.

9 is a flowchart illustrating a credit risk management method through credit risk building according to the present invention.

Claims (14)

A raw data database that stores data on the financial statements of the company or individual for credit rating; A data management unit for selecting data from the raw data database to remove missing values of the selected data and managing data from which the missing values are removed; An analysis database storage unit for storing data of the data management unit; A model construction unit for selecting a variable of the credit rating model from the data of the analysis database storage unit and constructing a credit rating model from the data by logistic regression analysis from the selected variable; An analysis result storage database unit for storing analysis results of the credit rating model constructed from the model building unit; And A tool kit for building a credit risk model comprising a model evaluator for evaluating the creditworthiness of the company or individual using the analysis result storage database. The method of claim 1, The data management unit, In the selected data, an independent variable that is a variable that affects the credit rating and a response variable that is a variable representing a result of the credit rating according to the independent variable are designated, and based on the independent variable and the response variable, the basic statistics of the selected data A data definition module for calculating information; A data refining module for refining the selected data by detecting missing values in the independent variable; A data replacement module for replacing the missing value detected by the data purification module with replacement data; A basic statistical analysis module analyzing basic statistics based on the data purified by the data purification module; A correlation analysis module extracting correlation independent variables related to the credit rating by analyzing correlations between the independent variables based on the basic statistics analyzed from the basic statistical analysis module; And And a regression analysis module for verifying statistical significance with the response variable using the correlation independent variable extracted from the correlation analysis module. The method of claim 2, The model construction unit, Segmentation module for classifying the enterprise or individual by a predefined group using the data stored in the analysis database storage unit, Selecting an independent variable required for the credit rating using the data stored in the analysis database storage unit An independent variable selection module, and a logistic model module for extracting a cutoff point for credit evaluation using the independent variable to predict a credit level. The method of claim 3, wherein The model building unit further comprises a calibration module for performing a regression analysis of the credit rating predicted by the logistic model module using a nonparametric regression model. The method of claim 1, The toolkit for credit risk model building, characterized in that the data stored in the raw material database, the analysis database unit and the analysis result storage database unit is stored in an XML document format. Selecting data to be analyzed from an original data database in which data calculated for corporate or individual credit rating is stored; In the selected data, an independent variable that is a variable that affects the credit rating and a response variable that is a variable representing a result of the credit rating according to the independent variable are designated, and based on the independent variable and the response variable, the basic statistics of the selected data Calculating information; Purifying the data by detecting missing values in the independent variable among the selected data; Selecting independent variables for performing the credit evaluation by performing basic statistics, correlations between the independent variables, and regression analysis based on the purified data; Constructing a prediction table related to credit rating using the selected independent variable, and constructing a credit rating model using the configured prediction table according to logistic regression analysis; And Estimating the credit rating by calculating the credit rating value using the constructed credit rating model. The method of claim 6, Purifying the data, Determining whether a missing value exists in the selected independent variable, and if the missing value exists, replacing the missing value with replacement data; The substitute data is credit risk management method through a credit risk model, characterized in that the median or average value in the independent variable including the missing value. The method of claim 6, The credit rating model comprises a logistic model credit risk management method through the construction of a credit risk model. The method of claim 6, Estimating the credit rating, And re-estimating the credit rating using a nonparametric regression model and evaluating the credit using the reestimated credit rating. The method of claim 6, Building a credit rating model using the configured prediction table, Extracting cutoff points using the configured prediction table, and building a credit rating model according to the extracted cutoff points. The method of claim 6, The basic statistics are A credit risk management method, comprising a credit risk model, comprising the type of data, the role of each variable, the size of the data sample, the number of missing values, the mean, the variance, and the five-digit summary value of the data. The method of claim 6, The credit rating model is Credit risk management method through the establishment of a credit risk model, characterized in that the bankruptcy model of the enterprise or individual, customer purchase model of a particular financial product, card delinquency model of the card company or the insurance against the departure of the insured model. The method of claim 6, Building a credit rating model using the configured prediction table, Credit risk management method through the construction of a credit risk model, characterized in that further comprising the step of building a credit rating model using the regression analysis using the selected independent variable. A recording medium stored in a computer for executing the method of any one of claims 6 to 13.

Images