KR102330423B1 - Online default forecasting system using image recognition deep learning algorithm - Google Patents

Abstract

An online default prediction system using an image recognition deep learning algorithm collects financial statement information of a bankrupt company, trains a deep learning default prediction model using the financial statement information of the defaulted company, and uses the deep learning default prediction model In this way, it is possible to provide information on forecasting defaults on the financial statement information of the client company.

Description

ONLINE DEFAULT FORECASTING SYSTEM USING IMAGE RECOGNITION DEEP LEARNING ALGORITHM}

The description below relates to a technology that provides predictability of a company's default based on financial statements provided online.

Since the 1960s, default prediction techniques using accounting information have been continuously tried. First, William Beaver published Profile Analysis in 1966. Bieber's study defined the period from 1954 to 1964 as the research period, and defined bankruptcy, non-payment of principal of bonds, bankruptcy, and non-payment of preferred stock dividends as bankruptcy. Through this, a model for predicting bankruptcy was established in the form of identifying the financial ratio with a statistically significant difference between the bankrupt company and the surviving company. Bieber's study is a univariate model that uses financial statement ratios one at a time to predict a company's bankruptcy.

However, univariate studies cannot improve the accuracy of prediction, and it is difficult to explain the corporate bankruptcy phenomenon that appears as a complex effect among various variables. In other words, univariate analysis does not provide a means to measure or combine the relative importance of individual financial statement ratios in predicting bankruptcy. For example, a univariate analysis does not provide any results as to which company has a higher bankruptcy risk, a company with a high current and debt ratio and a company with a low current and debt ratio.

After that, a full-scale multivariate model was developed, starting with Alzheimer's z-score, a type of multivariate discriminant analysis. This is a kind of discriminant analysis, and his model uses five financial ratios, weights are given to each ratio to maximize the explanatory power of the model, and an overall discriminant score called Z score is calculated. Despite these advantages, multivariate analysis has the following limitations. First, as with univariate analysis, we cannot be sure that multivariate analysis includes all relevant financial statement ratios related to bankruptcy. Second, it is necessary to subjectively judge the score that can distinguish between a bankrupt company and a non-bankrupt company. Third, it cannot be applied to unlisted companies because the market value of the shareholder's shares is unknown.

The direction of research for predicting bankruptcy is developing to use the logit model during the 1980s and early 1990s. Unlike multivariate analysis, logit analysis has the advantage of being easier to understand because it expresses the probability of bankruptcy as a probability. The difference from Beaver's research in the 1960s is that various variables are used as independent variables. Logistic regression analysis, a type of multiple regression analysis, uses the logit regression equation, a regression analysis method used when the dependent variable is a qualitative variable (bivariate) that appears in two response categories, and converts the calculated logit into probability. In the case of conversion, if the value exceeds 0.5, it is classified as a bankrupt company.

Recently, ensemble techniques such as random forest, which predict default by majority vote by generating a plurality of such logit models and then accumulating default or survival judgments of each model, are mainstream. Among the techniques developed so far, it has the highest default prediction rate.

Summarizing the above, default prediction techniques have been developed in two major directions: increasing the amount of independent variables and developing new algorithms. However, even the random forest technique, which has recently been applied in many cases, is actually limited in the amount of independent variables used in each model. This is because it is difficult to definitively determine which factors among a company's financial ratio led to bankruptcy in an attempt to use all available variables for default prediction. Therefore, it is necessary to take an innovative approach to predict the occurrence of bankruptcy due to the complex interaction between them by considering all variables.

The present invention includes the financial ratio and non-financial information (macroeconomic indicators, stock price-related indicators, accounting audits, other external evaluation indicators, etc.) as independent variables, and recognizes their multi-year changes as image patterns to reduce the possibility of long-term and short-term default. It provides a method and procedure for configuring the deep learning technique used for image recognition as an algorithm to determine default according to default prediction, and a model that can provide it as an online service.

A computer system comprising: at least one processor implemented to execute computer readable instructions contained in a memory, the at least one processor comprising: collecting financial statement information of a bankrupt company; training a deep learning default prediction model using the financial statement information of the defaulted company; and using the deep learning default prediction model to provide a computer system for processing the process of providing default prediction information for the financial statement information of the client company.

According to one aspect, the at least one processor may collect the financial statement information of the bankrupt company from the database system in conjunction with the database system of the institution holding the financial statement information of the company.

According to another aspect, the at least one processor may pre-process the financial statement information of the bankrupt company into image pixel information on a two-dimensional plane recognizable in a deep learning model for image recognition. In addition, in addition to financial statement information, macroeconomic indicators related to corporate activities, stock price indicators, accounting audit or other external evaluation indicators, etc. can be added to each pixel of the 2D image data as separate channel information. . This is similar to the principle that even if it is a two-dimensional image, one pixel has channel information having three values, such as RGB, so that it actually has three-dimensional information.

According to another aspect, the at least one processor is configured to be accurate not only in predicting the probability of default after one year, but also in the mid- to long-term forecast of two years or more, using the annual financial statement information or the financial statement information continuous for at least two years or more. This improved default prediction engine can be created.

According to another aspect, the at least one processor receives the financial statement information for each year of the defaulting company as an input and outputs labeling for at least one of default status, default occurrence time, and default type to dip You can train a learning model.

According to another aspect, the at least one processor may generate some missing financial statement information of the defaulted company using a generative adversarial network (GAN) model. Using the GAN model used to create a virtual image, even if there is some omission in the financial ratio information of the company subject to default prediction, it is possible to create the missing part according to the learned content. Through this, it is possible to predict default even if some missing financial statements are input or insufficient financial ratio information is input. can promote By using the GAN model, it is possible to create the entire financial ratio of a virtual company similar to a bankrupt company, rather than just restoring some missing financial ratio information. This virtual bankruptcy company financial ratio information is input to the CNN-based deep learning model and used for learning the default prediction engine. Since the financial ratio information of a hypothetical company is similar to that of a bankrupt company, but not an actual bankrupt company, CNN learns not to classify it as a bankrupt company. As such, the GAN and CNN set the loss function so that they are in an antagonistic relationship that should not deceive or deceive each other. can be expected

And, it is possible to generate a model for estimating default prediction through an arithmetic process between latent variables in a latent space using the GAN model. In the latent space of data where learning has been completed, the characteristics of GAN that can obtain a newly synthesized latent variable vector through various computational processes between latent variables are used for default prediction. For example, if there is a latent variable vector for predicting bankruptcy of small and medium-sized enterprises (SMEs) that have been trained, it is possible to derive a proxy model that can estimate the default of startups in the form of removing general characteristics of small and medium-sized enterprises and adding general characteristics of startups. can This can be used as a model for estimating the default rate when it is impossible to predict the default rate because default data has not yet been accumulated, such as a new startup.

According to another aspect, the at least one processor may modify a deep learning model parameter and a hyper parameter of the deep learning default prediction model using a recurrent neural network (RNN) model.

First, high default forecasting accuracy can be expected by directly using the entire financial ratio index of a company to predict default.

Second, it continuously collects financial ratio indicators from the birth of a company and recognizes the changes in financial ratio indicators within the entire life cycle as if it is a change in image, and informs whether deep learning corresponds to a company expected to default, so Early warning is possible. This can be expected to have superior capabilities compared to the forecast period of the current default forecasting system, which can predict only just before bankruptcy. This is because the algorithm is applied to analyze the image itself by patterning the influence relationship between various indicators by inputting the entire financial statement information rather than artificially selecting or shortening the information. For example, in the case of a bankrupt company's financial ratio, input data even if there is a unique financial ratio that appears only in a specific year and does not appear thereafter, or if there is a unique pattern in the financial ratio that occurs long before the time of bankruptcy. It is not a problem at all in image recognition deep learning that judges the whole as one continuous image and discovers patterns.

Third, it is possible to predict bankruptcy for companies with only partial financial ratio information, such as startups and unlisted companies. For example, in the case of a generative adversarial network (GAN), which is one of the image recognition techniques, it is possible to generate and combine some missing image pixels, so it can be used for the detection of a fraud prediction algorithm.

In this way, the scope of default forecasting can be increased by increasing the accuracy of corporate default forecasting, extending the forecast period to enable early warning, and extending the default forecast target companies to companies that have partially secured financial ratio indicators such as startups and unlisted companies. Due to the advantage of being able to do so, the utility of default forecasting can be greatly increased.

In particular, in the case of start-ups or unlisted companies, it is difficult to obtain adequate data for predicting bankruptcy, so loans are often rejected by these companies in the loan review despite the company's growth potential. Therefore, with this patent, which can predict default even under partial financial information, it can be expected to predict default for these companies, approve loans, and even vitalize startups.

In this way, if startups existing in the current market participate in online evaluation of a certain number of samples or more, the inconvenience of having to find the actual default rate when the company actually goes bankrupt, and the default rate can be calculated immediately at this point.

1 is a block diagram illustrating an example of an internal configuration of a computer system according to an embodiment of the present invention.
2 is a block diagram for explaining the internal configuration of an online default prediction system using an image recognition deep learning algorithm according to an embodiment of the present invention.
3 is an exemplary diagram for explaining the similarity between image pixels and financial ratio information according to an embodiment of the present invention.
4 is an exemplary view for explaining the principle of restoring the missing part of financial statement information using GAN in an embodiment of the present invention.
5 to 9 are exemplary views for explaining the default prediction training process in an embodiment of the present invention.
10 is an exemplary diagram for explaining a process of providing a default prediction result using a deep learning model in an embodiment of the present invention.

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

Embodiments of the present invention relate to a technology for providing predictability of bankruptcy based on its financial statements provided by a company online. It is to provide predictability of bankruptcy based on its financial statements provided by companies online, and for this purpose, it is a technology that applies artificial intelligence technology used for image recognition. Deep learning technology, which is one of the artificial intelligence techniques mainly used for image recognition, is a method of self-learning a mechanism for classifying images through a large number of input data. Conventionally, this has been utilized for image recognition, but in this embodiment, a method that can be applied to predict bankruptcy of a company is provided online.

In this embodiment, a method and procedure for introducing a deep learning technique used for image recognition as an algorithm that AI recognizes patterns between them while encompassing all financial ratio independent variables, and a model that can be used online are provided.

The basic form of the deep learning algorithm used for image recognition includes the process of extracting information in the form of repeating abstraction by taking into account all pixel values, partially abstracting information to summarize information, and summing it again. . Based on this abstraction, the key factors determining default are extracted from the company's financial statements, and default forecast information is provided online in real time using IT infrastructure and wired/wireless devices. To this end, the company's annual financial statements are stored for many years, generated as two-dimensional time series information, and used as image pixel information on a two-dimensional plane recognizable by an image recognition deep learning algorithm. In addition, non-financial information (macroeconomic indicators, stock price indicators, accounting audits, other external evaluation indicators, etc.) is also stored as time series information and used in the image recognition deep learning algorithm in the form of channel information added to each image pixel. Through this, the image recognition deep learning model is optimized for corporate bankruptcy detection.

In this embodiment, the following solutions are presented in order to overcome the limitations of corporate bankruptcy prediction, which remains in the standardized judgment criteria and short-term prediction.

First, it enables mid- to long-term forecasting of bankruptcy in corporate bankruptcy, secondly, it improves the accuracy of default forecasting, and thirdly, in the case of individual entrepreneurs and startups other than listed companies, it is customized according to the characteristics of various industries and corporate sizes, etc. We want to develop a detailed default prediction model. Fourth, when a mid- to long-term default forecast is observed, it is used in reverse to provide preventive consulting information to the relevant company so as not to follow the pattern of the defaulted company. It aims to supply related information to customers online by building the above services as IT-based mobile and PC solutions.

1 is a block diagram illustrating an example of an internal configuration of a computer system according to an embodiment of the present invention. For example, an online default prediction system according to embodiments of the present invention may be implemented through the computer system 100 of FIG. 1 . As shown in FIG. 1 , the computer system 100 is a component for executing the online default prediction method, and includes a processor 110 , a memory 120 , a persistent storage device 130 , a bus 140 , an input/output interface ( 150 ), and a network interface 160 .

Processor 110 may include or be part of any apparatus capable of processing a sequence of instructions as a component for online default prediction. Processor 110 may include, for example, a computer processor, a processor in a mobile device, or other electronic device and/or a digital processor. The processor 110 may be included in, for example, a server computing device, a server computer, a set of server computers, a server farm, a cloud computer, a content platform, and the like. The processor 110 may be connected to the memory 120 through the bus 140 .

Memory 120 may include volatile memory, permanent, virtual, or other memory for storing information used by or output by computer system 100 . The memory 120 may include, for example, random access memory (RAM) and/or dynamic RAM (DRAM). Memory 120 may be used to store any information, such as state information of computer system 100 . Memory 120 may also be used to store instructions of computer system 100 including, for example, instructions for online default prediction. Computer system 100 may include one or more processors 110 as needed or appropriate.

Bus 140 may include a communications infrastructure that enables interaction between various components of computer system 100 . Bus 140 may carry data between, for example, components of computer system 100 , such as between processor 110 and memory 120 . Bus 140 may include wireless and/or wired communication media between components of computer system 100 and may include parallel, serial, or other topological arrangements.

Persistent storage 130 is a component, such as memory or other persistent storage, as used by computer system 100 to store data for an extended period of time (eg, compared to memory 120 ). may include Persistent storage 130 may include non-volatile main memory as used by processor 110 in computer system 100 . Persistent storage 130 may include, for example, flash memory, a hard disk, an optical disk, or other computer readable medium.

The input/output interface 150 may include interfaces to a keyboard, mouse, voice command input, display, or other input or output device. Configuration commands and/or input for online default prediction may be received via input/output interface 150 .

Network interface 160 may include one or more interfaces to networks such as a local area network or the Internet. Network interface 160 may include interfaces for wired or wireless connections. Configuration commands and/or input for online default prediction may be received via network interface 160 .

Also, in other embodiments, computer system 100 may include more components than those of FIG. 1 . However, there is no need to clearly show most of the prior art components. For example, the computer system 100 is implemented to include at least some of the input/output devices connected to the above-described input/output interface 150, or a transceiver, a global positioning system (GPS) module, a camera, various sensors, It may further include other components such as a database and the like.

Hereinafter, a specific embodiment of an online default prediction system using an image recognition deep learning algorithm will be described.

2 is a block diagram for explaining the internal configuration of an online default prediction system using an image recognition deep learning algorithm according to an embodiment of the present invention. FIG. 2 is a diagram illustrating an example of components that the processor 110 of the computer system 100 in which the online default prediction system is implemented may include.

As shown in FIG. 2 , the processor 110 may include a collection module 201 , a preprocessing module 202 , a training module 203 , a service module 204 , and a feedback module 205 . These components of the processor 110 may be representations of different functions performed by the processor 110 according to a control instruction provided by at least one program code. For example, the collection module 201 may be used as a functional representation operative to control the computer system 100 such that the processor 110 collects financial statement information.

The processor 110 and components of the processor 110 may perform steps S310 to S330 included in the online default prediction method of FIG. 3 . For example, the processor 110 and components of the processor 110 may be implemented to execute an operating system code included in the memory 120 and an instruction according to at least one program code described above. Here, at least one program code may correspond to a code of a program implemented to process the online default prediction method.

The collection module 201 may collect financial statement information of a company or individual business operator online. The collection module 201 may collect financial statement information including financial statements or financial ratios of corporations or individual operators. The collection module 201 may collect the financial statement information of the bankrupt company from the database system of the institution by interworking with the database system of the institution holding the financial statement information of the corporation.

The pre-processing module 202 may pre-process the financial statement information collected by the collection module 201 into a form usable in various deep learning models for image recognition. The pre-processing module 202 may convert the financial statements or financial ratios of each company into two-dimensional time series information to be recognized as an image. The RGB color values of image pixels used in deep learning have the same maximum and minimum values. Therefore, it is converted by normalizing it to the same value in consideration of the variation range of each financial ratio indicator. This corresponds to a kind of pre-processing work in deep learning-based default prediction that replaces RGB values of image pixels with financial ratio index values. In addition, the preprocessing module 202 also plays a role in dividing data into a training set, a validation set, and a test set for model training and selection.

The training module 203 trains a deep learning model for default prediction using the financial statement information preprocessed in the preprocessing module 202, and the model parameters and hyperparameters of various deep learning models can be modified to increase the default prediction rate. have. The training module 203 may be in charge of predicting default by training the deep learning algorithm to correct the deep learning model parameters and hyperparameters using the preprocessed financial statement information.

In deep learning models such as CNNs, various parameters exist to increase the accuracy of prediction. By using the training set, the parameters of the deep learning model can be modified, and ultimately the predictability of the model can be increased. In addition, the actual predictive performance of the trained model can be verified by using the verification set. First, it uses CNN to extend the prediction period of bankruptcy and helps prevent it. To do this, the multi-year financial statements are input, and the labeling of whether or not default occurs, the expected time of default, and various default types (eg, lack of liquidity, capital erosion, etc.) Accordingly, as a prediction result, not only simple default, but also a probability of a default prediction timing or cause can be provided. Second, the discriminant model (module D) using GAN increases the predictive rate of default, and the generative model (module G) is not a simple criterion, but a company type that is not in the training set (e.g., new Start-up) can be inferred. In addition, if the default rate is investigated by classifying the training set according to various criteria, it is possible to estimate the probability of default for various types of companies by combining various criteria. Therefore, the default rate can be predicted with a more detailed standard, and through this, it is possible to predict the default rate for early start-ups where it is difficult to calculate the default rate because there is no default data right away, or even if there is data, it is possible to predict the default rate based on more various criteria. It is possible to improve the screening ability of companies in the loan system, etc. In particular, SGAN and ACGAN can be learned not only for default/survival discrimination, but also as an additional one-hot vector type, which can be subdivided into types of default or company size. can be higher

GAN can also create a new engine through computation in the latent space. As shown in [Potential space of defaulted company financial ratio - potential space of general corporate financial ratio + potential space of general startup financial ratio = potential space of defaulted startup financial ratio], even if default data is not accumulated for startups, learning is It is possible to produce typical data of completed startup bankruptcy (reduced to the so-called latent space).

Model selection can be implemented through hyperparameter tuning. A hyperparameter is a parameter that determines the model structure such as the number of layers in a deep learning model, and it is to find an optimal value according to the type and characteristics of data. Therefore, in this embodiment, the deep learning model can be selected by determining the hyperparameter using the test set.

In particular, when financial statements are partially provided due to partial loss or failure to prepare, or when erroneous financial statements exist due to errors or falsifications, original data can be generated by inferring them. For example, in the case of GAN, it includes a function to restore a damaged image or a mosaic-processed image to the original image, so it can be used for the creation and correction of financial statements.

The service module 204 receives financial statement information from the customer, applies the customer's financial statement information to various deep learning models trained in the training module 203, determines whether the customer defaults according to the financial statement information, and determines the result of the determination In accordance with the default forecast (prevention) information can be notified. The service module 204 may provide an online service for predicting default by using an IT solution such as a mobile or PC for online customers. In this case, the service module 204 may generate and provide consulting information to help the business management activities so that the online customer, the client company, does not follow the pattern of the bankrupt company.

The feedback module 205 may check the accuracy of default prediction by securing actual default occurrence data of the company later, and may modify the collection process, pre-processing process, and training process to improve the default forecasting accuracy based on the inspection result. . The feedback module 205 serves to modify the parameters by reflecting the actual corporate default results. For example, using a recurrent neural network (RNN) model to memorize model parameters and hyperparameters at a previous point in time. By modifying hyperparameters, an automated mechanism can be implemented that can reflect changes in actual default patterns within the changed corporate environment.

3 is an exemplary diagram for explaining the similarity between image pixels and financial ratio information.

Referring to FIG. 3 , as an example of a deep learning model for predicting bankruptcy, which is a training target in the training module 203 , the financial ratio information for each year is recognized like an image pixel value by a convolution neural network (CNN), and the information is abbreviated. can

As empirical analysis data on the lifespan determinants of all about 20,000 companies recorded in COMPUSTAT, the data of US listed companies, the financial ratio indicators that determine the lifespan of a company were analyzed over a period of about 40 years (1095~2015). It has a common influence on the stars, and this structure is stable and unchanging. This shows that the financial ratio indicator is stable regardless of the long or short life span of the company. In other words, it can be inferred that the financial ratio indicators can be used as an indicator of corporate default regardless of the age of the company.

The previous bankruptcy forecasts the possibility of the next default based on the previous forecast. Various algorithms can increase the default detection rate, but long-term forecasting is still a challenge. Thus, knowing the forecast of long-term default risk allows action to be taken before actual default occurs.

Various labeling is required depending on the cause of default. For example, both a surplus and a deficit. It is necessary to divide the company into large/small/medium companies and additionally classify them in detail, such as industry, but this can be solved with a generative adversarial network (GAN) model.

In addition, in order to predict the possibility of bankruptcy in the mid- to long-term before the point of bankruptcy, it is necessary to label the financial ratio image of the bankrupt company by classifying it from 1 year to n years of founding. have.

When using a deep convolutional generative adversarial network (DCGAN) as an example of a GAN model for default prediction:

1. Since the output of DCGAN is a latent space, the latent space can be obtained for each labeling. By appropriately adding or subtracting the latent space between each of these labeling, it is possible to create a new latent space and generate an image.

2. Using the principle of the GAN model, it is the same as <image of financial ratio of (listed company) insolvent company - image of financial ratio of surviving company = image of bankruptcy>, <image of bankruptcy + (unlisted company) financial ratio of insolvent company> Image = Same as image of unlisted company's financial ratio>. It is a form of determining how similar the financial ratio images of unlisted companies that have requested default predictions are, and can be substituted for default predictions of unlisted companies.

3. When applied to a small company that is most similar to a startup, the default image of the small company can be extracted, and if the financial ratio image of the startup is added to this, the prediction of the default of the startup is also possible as a proxy.

4. A business model can be built by receiving data from banks and institutions that have unlisted financial ratio data using IT infrastructure, updating them in real time, and transmitting the results to banks, institutions, and companies.

5. In the case of start-up companies, there is an advantage that some missing financial information can be generated using GAN and used as input data for default prediction. The principle of restoring the missing part of financial statement information using GAN is shown in FIG. 4 .

In the case of startup companies, the bankruptcy of startups can be predicted by inputting financial statements, condensing them with CNN, and comparing how similar they are to existing bankruptcy data.

The U.S. public company DB (COMPUSTAT) is stored for 45 years, so the lifespan of a bankrupt company is between 1 and 45 years. Therefore, since one deep learning model is required for each year, a total of 1,035 deep learning models are needed. It is trained in advance, matches the deep learning model to the financial statement data presented by the customer, and immediately informs the result. For example, if a customer has entered five years' worth of financial statements, find a deep learning model with a cluster size of 1 to 5 years, and read the customer's financial statements for a minimum of 1 year and a maximum of 5 years in a row. Enter it and report the default prediction result.

5 to 9 are exemplary views for explaining the default prediction training process.

5 to 7 show the process of learning the default prediction deep learning model by inputting the annual financial ratio with a cluster size of 1.

As shown in FIG. 5 , the collection module 201 collects annual financial ratio information for each company for each bankrupt company by year. The collection module 201 may collect information on the annual financial ratio of each company group by dividing the company group according to the company lifespan.

Referring to FIG. 6 , the training module 203 can generate a default prediction engine (D) by learning a default prediction deep learning model after one year using the financial ratio information for the last year of the entire company based on the time of default. have.

Similarly, referring to FIG. 7 , the training module 203 generates a default prediction engine (D) by learning a default prediction deep learning model after 2 years using the financial ratio information of the entire company two years ago based on the time of default. can

By repeating the above process, up to N default prediction engines (D) can be generated, but at least some of the default prediction engines (D), for example, default prediction engines (D) after a predetermined time based on the default time It can be excluded from the valid model target.

The above has been described as learning the deep learning model using annual financial ratio information, but it is not limited thereto, and it is also possible to learn the deep learning model using financial ratio information continuous for at least two years or more.

Referring to FIGS. 8 and 9 , the training module 203 may learn the default prediction deep learning model by using the financial ratio information for two consecutive years with a cluster size of 2, that is, for two years. As shown in FIG. 9, the training module 203 uses the financial ratio information for the last two years of the entire company based on the time of default by learning the deep learning model for predicting default after one year, thereby predicting default engine (D) can create Similarly, the training module 203 generates the default prediction engine (D) by learning the default prediction deep learning model after 2 years using the financial ratio information 2 years ago and the financial ratio information 3 years ago of the entire company based on the time of bankruptcy. can

Accordingly, the training module 203 can learn the default prediction deep learning model using successive financial ratios such as 2-year, 3-year, and 4-year units as well as 1-year financial ratios, and through this, bankruptcy prediction deep learning model can be learned. A full set of deep learning models for prediction can be constructed.

10 is an exemplary diagram for explaining a process of predicting default by using a deep learning model.

Referring to FIG. 10 , the service module 204 selects a deep learning model applicable to the customer from among the entire set of deep learning models generated by the training module 203 when receiving financial statement information from the customer (company). The bankruptcy point can be determined through the selected deep learning model. When a customer enters financial statement information for two years, it selects a deep learning model created with annual financial ratios for each year’s financial statement information, as well as selects a two-year financial ratio for two-year financial statement information. Deep learning models can be selected. The service module 204 may provide default prediction information corresponding to the financial statement information of the client customer as a service result.

As such, according to the embodiments of the present invention, high default prediction accuracy can be expected by directly using the entire financial ratio index of a company for default prediction. According to embodiments of the present invention, deep learning corresponds to a company expected to default by continuously collecting financial ratio indicators from the birth of a company, and recognizing changes in financial ratio indicators shown within the entire life cycle as if it were a change in image. It is possible to provide an early warning of default forecasting because it informs you whether According to embodiments of the present invention, default prediction is also possible for a company in which only some financial ratio information exists, such as startups and unlisted companies.

The device described above may be implemented as a hardware component, a software component, and/or a combination of the hardware component and the software component. For example, the devices and components described in the embodiments may include a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, a field programmable gate array (FPGA), and a programmable logic unit (PLU). It may be implemented using one or more general purpose or special purpose computers, such as a logic unit, microprocessor, or any other device capable of executing and responding to instructions. The processing device may execute an operating system (OS) and one or more software applications executed on the operating system. The processing device may also access, store, manipulate, process, and generate data in response to execution of the software. For convenience of understanding, although one processing device is sometimes described as being used, one of ordinary skill in the art will recognize that the processing device includes a plurality of processing elements and/or a plurality of types of processing elements. It can be seen that can include For example, the processing device may include a plurality of processors or one processor and one controller. Other processing configurations are also possible, such as parallel processors.

The software may comprise a computer program, code, instructions, or a combination of one or more thereof, which configures a processing device to operate as desired or is independently or collectively processed You can command the device. The software and/or data may be embodied in any type of machine, component, physical device, computer storage medium or device for interpretation by or providing instructions or data to the processing device. have. The software may be distributed over networked computer systems, and stored or executed in a distributed manner. Software and data may be stored in one or more computer-readable recording media.

The method according to the embodiment may be implemented in the form of program instructions that can be executed through various computer means and recorded in a computer-readable medium. In this case, the medium may be to continuously store the program executable by the computer, or to temporarily store the program for execution or download. In addition, the medium may be a variety of recording means or storage means in the form of a single or several hardware combined, it is not limited to a medium directly connected to any computer system, and may exist distributed on a network. Examples of the medium include a hard disk, a magnetic medium such as a floppy disk and a magnetic tape, an optical recording medium such as CD-ROM and DVD, a magneto-optical medium such as a floppy disk, and those configured to store program instructions, including ROM, RAM, flash memory, and the like. In addition, examples of other media may include recording media or storage media managed by an app store for distributing applications, sites for supplying or distributing other various software, and servers.

As described above, although the embodiments have been described with reference to the limited embodiments and drawings, various modifications and variations are possible from the above description by those skilled in the art. For example, the described techniques are performed in a different order than the described method, and/or the described components of the system, structure, apparatus, circuit, etc. are combined or combined in a different form than the described method, or other components Or substituted or substituted by equivalents may achieve an appropriate result.

Therefore, other implementations, other embodiments, and equivalents to the claims are also within the scope of the following claims.

Claims (7)

In a computer system,
at least one processor implemented to execute computer readable instructions contained in a memory
including,
the at least one processor,
a collection module for collecting financial statement information of a bankrupt company;
a training module for training a deep learning default prediction model using the financial statement information of the defaulted company; and
A service module that provides default prediction information on the financial statement information of a client company using the deep learning default prediction model
including,
The collection module is
Collecting the financial statement information of the bankrupt company from the database system in conjunction with the database system of the institution holding the company's financial statement information,
For the above-mentioned bankrupt companies, annual financial statement information for each company is collected by year,
By dividing the company group according to the company lifespan, information on the annual financial statements of each company group is collected,
The training module is
A deep learning model CNN (convolution neural network) model is trained by inputting the financial statement information for each year of the defaulted company as an input and outputting labeling for default status, default occurrence time, and default type as output,
After obtaining a latent space for each labeling using a generative adversarial network (GAN) model, a model for estimating default prediction through an arithmetic process between latent variables in the latent space is generated,
Construct a set of deep learning models for default prediction by creating a default prediction engine by year based on the time of default by using annual financial statement information or financial statement information continuous for at least two years,
The service module is
Selecting a deep learning model corresponding to the financial statement information of the client company from the deep learning model set and determining the default point through the selected deep learning model,
the at least one processor,
When an actual default occurrence company is reported, a feedback module that checks the default prediction information of the corresponding company and corrects the deep learning model parameters and hyper parameters of the deep learning default prediction model using a recurrent neural network (RNN) model based on the check result
A computer system further comprising a. delete According to claim 1,
the at least one processor,
A preprocessing module for preprocessing the financial statement information of the bankrupt company into image pixel information on a two-dimensional plane that can be recognized by a deep learning model for image recognition
further comprising,
The pre-processing module is
Pre-processing by adding non-financial information including macroeconomic indicators, stock price-related indicators, accounting audits, and external evaluation indicators of the bankrupt company to each pixel of image data as separate channel information
A computer system characterized by a. delete delete According to claim 1,
The training module is
Restoring some missing financial statement information of the insolvent company using the GAN model
A computer system characterized by a. delete

Images