JP7216854B1 - Information processing device and program - Google Patents

Abstract

An object of the present invention is to make it possible to estimate the fact of fraud with high accuracy when there is fraud in accounting processing, and to objectively present the grounds for judging fraud. An information processing apparatus (10) has acquisition means (1002), extraction means (1004), and determination means (1006). Acquisition means 1002 acquires financial data relating to a plurality of accounting periods for a target business operator. The extraction unit 1004 extracts, for each account item generated based on the financial data acquired by the acquisition unit 1002, transition information indicating the transition of the value of the account item. Judgment means 1006 judges the inappropriateness of accounting processing by the target business operator based on the transition information extracted by extraction means 1004 . [Selection drawing] Fig. 2

Description

The present invention relates to an information processing apparatus for inspecting accounting processing.

The number of companies engaged in inappropriate accounting, such as the falsification of accounting documents, is increasing year by year. In addition, considering the impact of poor business performance caused by the spread of the new coronavirus, it is thought that the high risk of inappropriate accounting will continue in the future. Thus, the need for inspection of accounting documents is increasing year by year. On the other hand, the shortage of personnel involved in accounting audits, such as accountants, has become a problem. Under these circumstances, in recent years advanced data analysis technologies such as machine learning and deep learning have been widely used, and it has been proposed to introduce these data analysis technologies into financial analysis in accounting audits. (For example, Patent Document 1).

Japanese Patent No. 6345856

However, the current situation is that the accuracy of mechanically determining whether or not fraud has occurred in accounting processing is not sufficient. One of the reasons why the judgment accuracy does not improve is that only a small part of the companies commit fraud, and as a result, the financial data used as training data is the financial data that has been fraudulent (i.e., the originally desired data). Since only parts are included, this makes it difficult to improve the learning efficiency. In addition, when it is determined that there is fraud, there is a problem in presenting the grounds objectively. Determining the presence or absence of fraud is an important matter that can be a matter of life and death for the company. is important.

SUMMARY OF THE INVENTION It is an object of the present invention to provide a technique for estimating the existence of fraud in accounting processing with high accuracy and objectively presenting the grounds for judging fraud.

An information processing apparatus according to one aspect of the present invention includes acquisition means, extraction means, and determination means. The acquiring means acquires financial data relating to a plurality of accounting periods for one business operator. The extracting means extracts, for each account item generated based on the financial data acquired by the acquiring means, transition information indicating the transition of values related to the account item. The judging means judges the inappropriateness of the accounting processing by the one business operator based on the transition information extracted by the extracting means.
According to this information processing device, when there is fraud in accounting processing by one business operator, the fact of fraud can be estimated with high accuracy. In addition, since the basis for judging fraud can be presented objectively, subsequent manual verification becomes efficient.

In a preferred aspect, the determination means has preprocessing means, a plurality of weak learners, and bagging means. The preprocessing means performs undersampling based on the transition information extracted by the extracting means to generate a plurality of data sets. A plurality of data sets output from the preprocessing means is input to each of the plurality of weak learners. The plurality of weak learners construct a learning model for outputting the indicator of inappropriateness based on the financial data of the plurality of business operators. The bagging means calculates an index indicating inadequacy by bagging outputs from the plurality of weak learners.
According to this aspect, the fact of fraud can be estimated with higher accuracy. In addition, it is possible to provide a learning method suitable for a situation in which it is difficult to collect fraudulent training data.

In a preferred aspect, in the information processing apparatus according to the present invention, a plurality of weak learners operate based on gradient boosting.
According to this aspect, the accuracy of determining the presence or absence of fraud in accounting processing is improved.

In a preferred embodiment, the preprocessing means generates data by adding feature amounts obtained by inputting the financial data to a plurality of learning models different from the learning model to transition information extracted by the extraction means. . The preprocessing means then generates the plurality of data sets by undersampling the generated data.
According to this aspect, the accuracy of determining the presence or absence of fraud in accounting processing is improved.

In a preferred embodiment, the determination means calculates a SHAP value indicating the degree of contribution for each account based on the index calculated for each account.
According to this aspect, the basis for judging fraud is visualized.

In a preferred aspect, the preprocessing means corrects the transition information based on market data.
According to this aspect, the transition information, which serves as a reference for judging inappropriateness in accounting processing, is corrected based on the market condition data, so that judgment accuracy is improved.
In another aspect, the present invention provides a step of acquiring financial data relating to a plurality of accounting periods for one business operator in a computer; Provide a program for executing the steps of extracting transition information indicating the transition of values related to items, and determining inappropriateness in accounting processing by the one business operator based on the generated information. do.

1 is a diagram showing a configuration example of an information processing device 10 according to an embodiment of the present invention; FIG. 3 is a functional block diagram showing functions realized by a control unit 100 of the information processing device 10 according to a financial analysis program PA; FIG. 4 is a flow chart showing the flow of a financial analysis method executed by the control unit 100 of the information processing device 10 according to the financial analysis program PA. 4 is a diagram showing an example of SHAP values displayed on the display unit 130 of the information processing device 10; FIG.

Various technically preferable limitations are attached to each embodiment described below. However, embodiments of the present invention are not limited to the forms described below.

A. Embodiment FIG. 1 is a diagram showing a configuration example of an information processing apparatus 10 according to an embodiment of the present invention. The information processing device 10 is, for example, a computer device owned and managed by an audit corporation, but the administrator or user of the computer device does not matter. The information processing apparatus 10 in this embodiment is, for example, a personal computer. As shown in FIG. 1, the information processing apparatus 10 includes a control unit 100, a communication I/F unit 110, an operation input unit 120, a display unit 130, a storage unit 140, and a bus for transferring data between these components. 150.

The control unit 100 is, for example, a CPU (Central Processing Unit). The control unit 100 functions as a control center of the information processing device 10 by executing the financial analysis program PA stored in the storage unit 140 .

Communication I/F unit 110 is connected wirelessly or by wire to an electric communication line such as the Internet. Communication I/F section 110 receives data sent via an electric communication line and transfers the received data to control section 100 . Further, communication I/F section 110 transmits data given from control section 100 to an electric communication line.

Operation input unit 120 includes either one or both of a pointing device such as a mouse and a keyboard. When the user of the information processing apparatus 10 performs an operation on the operation input unit 120 , the operation input unit 120 outputs operation content data indicating the content of the user's operation to the control unit 100 . Thereby, the content of the user's operation is transmitted to the control unit 100 .

The display unit 130 is a display device including a liquid crystal panel and its driving circuit. The display unit 130 displays various images under the control of the control unit 100 .

Storage unit 140 is a storage device including volatile storage unit 142 and nonvolatile storage unit 144 . The volatile storage unit 142 is, for example, a RAM (Random Access Memory). The volatile storage unit 142 is used by the control unit 100 as a work area when executing various programs. The nonvolatile storage unit 144 is, for example, a hard disk. Various programs and various data are stored (installed) in advance in the nonvolatile storage unit 144 .

As an example of the programs stored in the nonvolatile storage unit 144, a kernel program for realizing an OS (Operating System) in the control unit 100 and a financial analysis program PA are stored in advance. Note that illustration of the kernel program is omitted in FIG. The financial analysis program PA is a program that causes the control unit 100 to execute the financial analysis method according to the present invention.

When the information processing apparatus 10 is powered on (not shown in FIG. 1), the control unit 100 reads the kernel program from the nonvolatile storage unit 144 to the volatile storage unit 142 and starts executing the kernel program. Control unit 100 operating according to the kernel program receives from operation input unit 120 operation content data for instructing the start of execution of another program. It reads out to the storage unit 142 and starts its execution. In the following, the case where the operation input unit 120 is operated to instruct the execution of the financial analysis program PA will be mainly described.

Upon receiving operation content data instructing start of execution of the financial analysis program PA from the operation input unit 120, the control unit 100 reads the financial analysis program PA from the nonvolatile storage unit 144 to the volatile storage unit 142 and starts its execution. do. The control unit 100 operating according to the financial analysis program PA implements the functions of the present invention.

FIG. 2 is a functional block diagram showing functions realized by the control unit 100 according to the financial analysis program PA. As shown in FIG. 2, the control unit 100 operating according to the financial analysis program PA functions as acquisition means 1002, extraction means 1004, and determination means 1006. That is, the acquiring means 1002, the extracting means 1004, and the determining means 1006 shown in FIG. 2 are software modules realized by operating a computer such as a CPU according to software (financial analysis program PA). The functions of the acquiring means 1002, the extracting means 1004, and the determining means 1006 are as follows.

Acquisition means 1002 acquires financial data D1 related to a plurality of accounting periods, such as five terms, for one business operator subject to risk management (hereinafter referred to as the target business operator). A business operator can be a corporation, other organization, sole proprietor, etc., as long as it conducts some kind of business activity and prepares and manages accounting information, and does not necessarily have to match the legal definition or classification. . So-called group companies, including consolidated subsidiaries, etc., may be regarded as a single business operator. In short, business entities related to transactions described in one financial data D1 can be regarded as one business operator.

Acquisition means 1002 acquires financial data D1 from a server in charge of financial management in the target business operator by communicating with the server via communication I/F section 110 . Financial data for one accounting period typically represents a financial statement that describes the values at the accounting period for each account item such as cost of sales, sales, and inventories. In other words, the financial data D1 represents financial statements for each of a plurality of accounting periods. However, the financial data D1 does not need to indicate the contents of the accounting book or account statement itself prepared by the business operator, and may be created based on the records such as the accounting statement or account book. For example, in the case where there are only accounts for sales and profits on the financial statements, the financial data D1 may include information generated using information on a plurality of account items such as profit margin on sales.

The financial data D1 may be in a format required by regulatory agencies such as the Financial Services Agency or by a legal institution such as a stock exchange, or may be in a format unique to the target company. may In short, the financial data D1 is a ledger in which the contents of past transactions performed by the business operator are described, and it is sufficient if it is composed of headings (account items, display items) and their contents. It doesn't matter if it's public or private. The financial data D1 may or may not be associated with information indicating whether or not fraud has actually been discovered.
In addition, the accounting period does not necessarily have to match the articles of incorporation of the business operator or required by law, and may be defined as the division of the above-mentioned account books.
Note that if the financial data D1 is open to the public by a supervisory authority or a stock exchange, the obtaining means 1002 may acquire the financial data D1 not from the server of the target business operator but from the supervisory authority or the stock exchange. good.

The extracting means 1004 extracts from the financial data D1 the transition information D2 indicating the transition of values over a plurality of accounting periods for each account item extracted based on the financial data D1 acquired by the acquiring means 1002 . For example, if M (M is an integer equal to or greater than 2) account items are included in the financial statement of the target business operator, the extracting means 1004 extracts M account items at maximum. When M account items are extracted, the transition information D2 represents the transition of values over a plurality of accounting periods for each of the M account items.
Note that the extraction means 1004 may store the extracted transition information D2 in the non-volatile storage unit 144 so that the grounds for judging fraud can be presented at a later date. Based on the stored transition information, the display unit 130 may display a graph or the like showing the transition of values for each account item over a plurality of accounting periods of the target business operator.

Judgment means 1006 judges the inappropriateness of the target business operator's accounting process based on the transition information D2 generated by the extraction means 1004 . As shown in FIG. 2, the determination means 1006 in this embodiment has a preprocessing means 1006a, weak learners 1006b(1) to 1006b(N), and a bagging means 1006c. Note that N is an integer of 2 or more. Below, weak learners 1006b(1) to 1006b(N) are referred to as weak learners 1006b when there is no need to distinguish between the weak learners 1006b(1) to 1006b(N).

As described above, the determination means 1006 has a plurality of weak learners 1006b. A weak learner is an AI that cannot expect high prediction accuracy by itself. A specific example of a weak learner is a decision tree. A decision tree is a tree structure for deriving a conclusion about the target value of a certain item from the observation results for that item. In this embodiment, a learning model (prediction model) for outputting an indicator indicating inappropriateness of accounting treatment for each account based on financial data of multiple businesses different from the target business is a weak learning model. constructed in the form of an ensemble of units 1006b(1) through 1006b(N).

More specifically, the weak learners 1006b(1) to 1006b(N) in this embodiment function as the learning model by operating based on gradient boosting. Gradient boosting is a method in which partial data is randomly obtained from the training data, a model (decision tree) is constructed sequentially, and the final predicted value is obtained by a weighted majority vote of the predicted values of each model. say.

The preprocessing means 1006a generates a plurality of data sets based on the transition information D2 extracted by the extraction means 1004. FIG. More specifically, the preprocessing means 1006a first inputs financial data to a plurality of learning models different from the learning models constructed by the weak learners 1006b(1) to 1006b(N), and obtains them. Data is generated by adding the obtained feature quantity to the transition information D2. Specific examples of a plurality of learning models in this embodiment include an isolation forest, a CFO modified Jones model, and the like. Next, the preprocessing means 1006a performs undersampling on the data obtained by adding the feature amount obtained by inputting the financial data D1 to the plurality of learning models to the transition information D2 to generate a plurality of data sets, that is, the data set D3(1) to D3(N) are generated. Undersampling refers to randomly extracting Y (a positive integer equal to or less than X) data from X (X is an integer equal to or greater than 2) data. In this embodiment, the preprocessing means 1006a performs N times of undersampling to generate data sets D3(1) to D3(N), but the data extracted in each undersampling partly overlap each other. may In undersampling, data sets D3(1) to D3(N) must contain all of the transition information D2 of fraud cases in the past years.

Next, the preprocessing means 1006a inputs the data sets D3(1) to D3(N) to the weak learners 1006b(1) to 1006b(N) one by one. Specifically, preprocessing means 1006a inputs data set D3(n) to weak learner 1006b(n).

The bagging means 1006c performs bagging (in this embodiment, weighted By voting by majority, an index V1 indicating inappropriateness is calculated for each account item. When M account items are extracted from the financial data D1, the bagging means 1006c calculates M indexes V1, one for each of the M account items. The judging means 1006 judges the inappropriateness of the accounting process for each target business operator and account item based on the index V1. For example, if the index calculated by the bagging unit 1006c exceeds a predetermined threshold, the determination unit 1006 determines that there is a high possibility that inappropriate accounting processing is being performed. Note that the determination means 1006 may store the data indicating the indicator V1 and the determination result based on the indicator V1 in the non-volatile storage unit 144, and display the indicator V1 and the determination result based on the indicator V1 on the display unit 130. may

Also, the control unit 100 operating according to the financial analysis program PA executes the financial analysis method according to the present invention. FIG. 3 is a flow chart showing the flow of this financial analysis method. As shown in FIG. 3, this financial analysis method includes an acquisition process SA110, an extraction process SA120, and a determination process SA130.

In acquisition processing SA110, control unit 100 functions as acquisition means 1002. FIG. In acquisition processing SA110, the control unit 100 acquires financial data relating to a plurality of accounting periods for the target business operator.

In extraction processing SA120 subsequent to acquisition processing SA110, control unit 100 functions as extraction means 1004. FIG. In the extraction process SA120, the control unit 100 generates transition information indicating the transition of the value of the account item over a plurality of accounting periods for each account item extracted based on the financial data acquired in the acquisition process SA110. do.

In determination processing SA130 subsequent to extraction processing SA120, control unit 100 functions as determination means 1006. FIG. In determination processing SA130, control unit 100 first performs extraction processing SA120 on data sets (that is, N data sets) to be input to each of weak learners 1006b(1) to weak learners 1006b(N). Generated based on the transition information generated in . Next, the control unit 100 inputs each of the N data sets to each of the weak learners 1006b(1) to 1006b(N) one by one. 1006b(N) is bagged to calculate an inappropriateness index for each account. FIG. 4 shows an example of the index calculated for each account item and the total value of the index (estimated value indicating the probability of fraud being presumed).

In the figure, the higher the SHAP value for each account item, the more similar the behavior to past fraud cases. In addition, the total value of the SHAP value is calculated as a total risk score between 0 and 1 by logit transformation, and the higher the value, the greater the possibility that the target business operator's accounting treatment is inappropriate. indicates From the figure, it can be seen that indicators related to inventories such as "inventory turnover period" and "inventory to net asset ratio" are highly inappropriate as factors that push up the total of the indicators. Therefore, an auditor, etc., who sees this result, is suspected of overstating inventories, and can decide to check these items intensively. In addition, overstating of inventories leads to overstating of profits, but in fact, in this example, it can be confirmed that both profit-related indicators and SHAP values are high, and it is possible to judge that there may be overstating of inventories. Justification is provided.

According to the above-described embodiment, by analyzing the transition of values for each account over a plurality of accounting periods by machine learning, an index indicating the inappropriateness of accounting processing is calculated for each account. For example, if the value of a particular account item has changed sharply (e.g., by more than a predetermined threshold year-on-year) in the most recent year, or conversely, if a particular account item appears unnatural in multiple accounting periods. If it is as constant as possible, it will be reflected in the SHAP value for that account.

Specifically, by combining ensemble learning (bagging) using undersampling and weak learning, the situation unique to this field is that only a small portion of the financial data used as training data is fraudulent. Also in, learning efficiency can be improved. As a result, the inappropriateness of accounting treatment can be estimated with high accuracy. In addition, since the grounds for judging fraud are presented as SHAP values for each account item, objectivity and reliability are improved.

Regarding the financial data D1, when an auditor or the like re-verifies the results of the system according to the present invention, compared to the case where the system according to the present invention is not used in advance, check points (which account items are checked preferentially should be performed), it is expected to improve the efficiency of the verification work.

B. Other Examples The embodiments described above may be modified as follows.
When generating a plurality of data sets based on the transition information extracted by the extraction unit 1004, the preprocessing unit 1006a corrects the transition information based on the market condition data, and corrects the transition information based on the corrected transition information. Multiple datasets may be generated. Market conditions refer to trading conditions in various markets such as stock markets and commodity markets.
Market data includes stock price index, domestic corporate goods price index, exchange rate against the US dollar (or the average value of the exchange rate), credit growth rate, cash currency (banknotes issued + money in circulation) and deposits in domestic banks, etc. data representing any one or more of the total deposited deposits, lending rates, and long-term bond yields. A stock price index is the year-on-year change rate of a stock price. The domestic corporate goods price index is the rate of increase or decrease in the total value added of domestically produced goods and services.

By reflecting the market conditions in the indicator that indicates the inappropriateness of accounting treatment, for example, in the financial data D1, if the sales have dropped significantly in the most recent year while the economy has deteriorated sharply in the past year , from the learning results, it can be concluded that this decline in sales was caused by the deterioration of the economy and cannot be said to be direct evidence of fraud. On the other hand, if the economy were to perform well over the past year, the results would show that the decline in sales was caused by fraudulent accounting practices such as the falsification of books and false declarations. is assumed.

The preprocessing means 1006a may correct the transition information based on the business type, business type, industry to which the business belongs, and other attributes instead of or in addition to the market condition data. Since the characteristics and trends of financial data may differ depending on the nature of the business, it is expected that the accuracy of judgment will improve.

The determining means 1006 may calculate a SHAP value indicating the degree of contribution (degree of contribution) of each account item to the inappropriateness based on the index calculated for each account item. In this case, the SHAP value may be stored in the nonvolatile storage unit 144 and displayed on the display unit 130 .

Acquisition means 1002, extraction means 1004, and determination means 1006 in the above embodiments are software modules, but any one, a plurality, or all of the acquisition means 1002, extraction means 1004, and determination means 1006 are ASICs. It may be a hardware module such as Even if one, more than one, or all of the acquiring means 1002, the extracting means 1004, and the determining means 1006 are hardware modules, the same effects as those of the above embodiments can be obtained.

In the above embodiment, the financial analysis program PA that causes a computer such as a CPU to execute the financial analysis method of the present invention is stored in advance in the storage unit 140 of the information processing apparatus 10 . However, the financial analysis program PA may be manufactured alone, or transferred (that is, provided) with or without charge. As a specific mode of providing the financial analysis program PA, a mode of writing the financial analysis program PA in a computer-readable recording medium such as a flash ROM and distributing it, or distribution by downloading via an electric communication line such as the Internet. The aspect to carry out is mentioned. By operating a general computer according to the financial analysis program PA distributed according to these aspects, it becomes possible to cause the computer to execute the financial analysis method of the present invention, and the same effect as the above embodiment can be obtained. .

In short, in the information processing system according to the present invention, the step of acquiring financial data related to a plurality of accounting periods for one business operator, and for each account item generated based on the acquired financial data, the account item and a step of determining the inappropriateness of accounting processing by the one business operator based on the generated information.

DESCRIPTION OF SYMBOLS 10... Information processing apparatus 100... Control part 110... Communication I/F part 120... Operation input part 130... Display part 140... Storage part 150... Bus 1002... Acquisition means 1004... Extraction means 1006 Determination means 1006a Preprocessing means 1006b, 1006b(1) to 1006b(N) Weak learning device 1006c Bagging means PA Financial analysis program.

Claims (5)

Acquisition means for acquiring financial data related to multiple accounting periods for one business operator;
an extracting means for extracting, for each account item generated based on the acquired financial data, transition information indicating the transition of values relating to the account item;
determining means for determining inappropriateness in accounting processing by the one business operator based on the generated information;
The determination means is
preprocessing means for performing undersampling based on the extracted transition information to generate a plurality of data sets;
a plurality of weak learners to which a plurality of data sets output from the preprocessing means are respectively input, for outputting the indicator indicating the inappropriateness based on the financial data of the plurality of business operators; a weak learner for building a learning model;
Bagging means for calculating an index indicating the inappropriateness by bagging outputs from the plurality of weak learners,
calculating a SHAP value indicating the degree of contribution for each account based on the index calculated for each account;
Information processing equipment. The information processing apparatus according to claim 1, wherein the preprocessing means corrects the transition information based on market data. The information processing apparatus according to claim 1 or 2, wherein the plurality of weak learners operate based on gradient boosting. The pretreatment means is
generating data by adding the feature amount obtained by inputting the financial data to a plurality of learning models different from the learning model to the extracted transition information, and undersampling the generated data; The information processing apparatus according to claim 3, wherein said plurality of data sets are generated by performing. to the computer,
obtaining financial data for a plurality of accounting periods for one business operator;
a step of extracting, for each account item generated based on the acquired financial data, transition information indicating the transition of values related to the account item;
Based on the generated information, a step of judging the inappropriateness of accounting processing by the one business operator, and
The determining step includes:
undersampling based on the extracted transition information to generate a plurality of data sets;
Build a learning model for outputting the indicator indicating the inappropriateness based on the financial data of the plurality of business operators, using a plurality of weak learners to which the generated plurality of data sets are respectively input. and
Bagging outputs from the plurality of weak learners to calculate an index indicating the inadequacy;
and calculating a SHAP value indicating a degree of contribution for each account based on the index calculated for each account.

Images