Method and device for detecting a financial or accounting error and / or fraud
The method addresses inefficiencies in detecting accounting errors and fraud by applying compliance rules and machine learning algorithms to optimize fraud detection, enhancing precision and reducing false positives, enabling automatic and regular identification within companies.
Patent Information
- Application Number
- PCT/IB2023/054610
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2023-05-03
- Publication Date
- 2024-11-07
AI Technical Summary
Current methods for detecting accounting or financial errors and fraud are inefficient, relying on external audits that are costly and only retrospective, and existing self-learning systems are prone to overfitting and underfitting due to data disparities and the complexity of accounting data, leading to high false positives and a need for improved automatic detection systems.
A computer-implemented method that applies predefined compliance rules to accounting data to identify errors, followed by anomaly detection using configurable tests, and then submits anomalies to a machine learning algorithm trained to detect fraud, optimizing detection performance and reducing false positives by pre-processing and enriching the training database.
This method enhances the precision of fraud detection, reduces false positives, and allows for automatic and regular identification of accounting errors and fraud, freeing companies from external audits and improving data quality and detection performance over time.
Smart Images

Figure IB2023054610_07112024_PF_FP_ABST
Abstract
Description
[0001] METHOD AND DEVICE FOR DETECTING ACCOUNTING OR FINANCIAL ERRORS AND / OR FRAUD
[0002] Technical field
[0003] The present invention relates to a computer-implemented method and device for detecting accounting or financial errors and / or fraud.
[0004] Prior art
[0005] Every business is exposed to the risk of fraud and / or accounting errors. In particular, crime in this area has increased with the digitalization of accounting and financial transactions. The rise of phishing scams, malware infections, and ghost websites can lead to massive losses for businesses that fall victim to them. Furthermore, it is important to detect these online banking frauds as quickly as possible, as it is very difficult to recover losses if the fraud is not discovered at the time it occurs.
[0006] Additionally, every business is subject to strict regulations depending on the tax legislation to which it is subject. For example, businesses subject to tax in France must keep computerized accounts in a specific format and alert them if fraud is detected.
[0007] Managing, monitoring, and analyzing all accounting data requires considerable effort and specific financial expertise. Traditionally, companies therefore call on specialized external consulting firms to conduct ad hoc audits to ensure the compliance of their accounting and potentially detect fraud. However, in addition to resulting in significant recurring costs, these audits only allow fraud to be identified after the fact. In addition, there is a dependency on external consulting firms.
[0008] W02017 / 210519 describes a self-learning system for isolating transactions detected as potentially fraudulent by applying a first algorithmic model identifying potential fraudulent transactions, then by applying a second model to reduce the number of false positives, i.e. reducing the number of transactions identified as fraudulent when they are not fraudulent. Such a system allows the user to be alerted in the event of suspected fraud being detected. However, this type of approach is particularly sensitive to overfitting (or "overfilling") and underfitting (or "underfilling"). These phenomena are notably due to the disparity of the data and the frequency of events to be detected in the data. Due to the large amount of accounting data and its variety, it is particularly complex and tedious to label and clean accounting and financial data.In particular, data that may appear suspicious in one context will not necessarily be so in another context.
[0009] US 2019 / 377624 describes a method for analyzing accounting data to detect errors and classify error types (aberrant errors, errors based on non-compliance with a rule and semantic errors).
[0010] US2019 / 377624 does not allow for the automatic detection of fraudulent data and requires an analysis, by an operator, of the data detected as erroneous in order to possibly detect fraudulent data.
[0011] Statement of the invention
[0012] There is a need to help businesses comply with regulations and enable them to automatically and regularly identify accounting or financial errors and any suspected fraud. In particular, there is a need to allow businesses to free themselves from ad hoc audits and external interventions.
[0013] Furthermore, in the context of fraud detection, it is essential that systems offer high accuracy, allowing the detection of any suspicious activity while limiting the number of false positives. There is therefore a need to improve automatic detection systems in order to achieve acceptable performance for the accounting and financial field.
[0014] Generally, there is a need to facilitate the monitoring of accounting and financial activities within companies.
[0015] The invention aims to meet these needs. Summary of the invention
[0016] The invention relates to a method, implemented by at least one computer, for detecting fraud from accounting data, the method comprising the following successive steps: a) applying a set of predefined compliance rules to the accounting data so as to identify erroneous accounting data, for each erroneous accounting data identified, pre-processing said erroneous accounting data; b) detecting accounting data constituting anomalies by applying a set of configurable tests to the accounting data from step a); c) detecting fraud by submitting the anomalies detected in step b) to a machine learning algorithm trained to detect, from input data consisting of accounting data anomalies, accounting data relating to fraud and possibly false positives.
[0017] A method according to the invention advantageously makes it possible to control the quality of the data through the application of conformity rules and pre-processing.
[0018] In particular, step a) of a method according to the invention makes it possible to attribute a degree of reliability to each piece of data, based on applicable compliance rules, regardless of the accounting data considered.
[0019] Furthermore, unlike known methods, a method according to the invention searches for fraudulent data among abnormal data. By revaluing the frequency of appearance of data considered fraudulent compared to the data provided as input to the machine learning algorithm, the method according to the invention allows better optimization of detection performance.
[0020] Finally, the method according to the invention also makes it possible to detect false positives and to improve detection performance over time.
[0021] Pre-processing of accounting data
[0022] Accounting data may include accounting entries made up of accounting lines, for example from at least one accounting entry file.
[0023] An erroneous accounting data may contain one or more errors, with the preprocessing depending on the error or errors. In one embodiment, the compliance rules are applied sequentially to the accounting data.
[0024] Alternatively, compliance rules are applied alongside accounting data.
[0025] In another embodiment, the compliance rules are partly applied sequentially and partly applied in parallel.
[0026] The sequential and / or parallel application of rules advantageously allows for the optimization of computing resources and processing time.
[0027] The pre-processing in step a) may in particular include deletion, assignment of a confidence index, or weighting of the accounting data considered.
[0028] A compliance rule can be a blocking rule, involving the deletion of accounting data during pre-processing if the accounting data does not comply with the blocking compliance rule.
[0029] Preferably, at least one compliance rule is a blocking rule involving the deletion of accounting data during pre-processing if said accounting data does not comply with the blocking rule, preferably less than 50% of the compliance rules are blocking, preferably even less than 25% of the compliance rules.
[0030] The method according to the invention does not necessarily involve the deletion of any data that does not comply with a compliance rule. Such systematic deletion could, on the contrary, be detrimental in the detection of fraudulent data, which could inadvertently be deleted before being detected.
[0031] In step a), the set of predefined compliance rules may include more than 5 compliance rules, preferably more than 10 rules, preferably more than 20 rules.
[0032] For example, the predefined compliance rule set includes at least one, preferably more than one, of the following compliance rules:
[0033] - verification of the information in mandatory fields to be completed in accounting data;
[0034] - verification of the format of fields to be filled in accounting data, in particular the format of dates;
[0035] - verification of the presence of a debit amount or a credit amount;
[0036] - checking for the absence of a debit amount and a credit amount simultaneously; - checking the currency code for amounts in foreign currency, for example, does euro correspond to the code “EUR” and / or the numeric code “978” of the ISO 4217 standard;
[0037] - verification of the language according to the regulations of the country to which the accounting data must comply, in particular for accounting data relating to French regulations, verification of the absence of words in English;
[0038] - at least one rule verifying the conformity of accounting data with a national general accounting plan;
[0039] - at least one rule verifying accounting data with a chart of accounts of a client from which the accounting data comes;
[0040] - for accounting data being an accounting entry file, verification of the chronology of the accounting entries and / or verification of the absence of sequence inversions and / or verification of the absence of sequence gaps and / or verification of the balance of the entries, in particular the balance of the entries according to the currencies and / or verification of the absence of currency other than euro with a zero amount in currency and / or verification of the dates according to the period concerned by the accounting entry file and / or verification of a lettering subsequent to the accounting entry,and in particular the verification of a lettering subsequent to the validation of the entry and / or verification that the date of an accounting entry is prior to the validation date of said accounting entry and / or verification that a validation period for an accounting entry is less than or equal to 90 days and / or verification of consistency between the different fields of a line of an accounting entry and / or verification of a framework with the general balance and / or verification that an accounting entry includes at least two accounting lines.,
[0041] Step a) is used to check that the accounting data complies with the standards applicable to the field of accounting and finance. This produces reliable accounting data that can be compared with each other. This pre-processing improves the performance of the detection algorithms, while limiting the loss of information contained in the accounting data. Anomaly detection
[0042] In step b), the accounting data from step a) are subjected to a set of configurable tests in order to detect among the accounting data the data constituting anomalies.
[0043] Anomalies can include a gross error, an acceptable error, a false positive, or fraud. A single event can constitute a false positive, for example, a one-off transfer to an unusual subcontractor for a relatively large amount can be detected as an anomaly.
[0044] At least one test from the set of configurable tests can be configured by modifying a threshold value, for example the threshold value being representative of an acceptable deviation with regard to a customer context.
[0045] Accounting data is considered an anomaly depending on the customer context to which the data belongs. The configuration of configurable tests thus advantageously makes it possible to adapt anomaly detection according to this context.
[0046] The set of parameterizable tests preferably includes at least 5 tests, preferably at least 10.
[0047] Preferably, the set of parameterizable tests includes at least one statistical test, verifying whether an accounting data is statistically distant or statistically close to the other accounting data considered. For example, the statistical test can be based on Benford's or Chebyshev's law or the normal law.
[0048] Preferably, the parameterizable test set is composed of more than 50 algorithms, preferably more than 100 algorithms, preferably more than 200 algorithms, more preferably more than 300 algorithms.
[0049] For example, the set of configurable tests includes all or part of the following tests:
[0050] - accounting data involving unusual accounts relative to a client context and / or relative to the accounting data resulting from steps a);
[0051] - accounting data involving rarely used accounts in relation to a client context and / or in relation to the accounting data from steps a);
[0052] - presence of abnormally high or low amounts relative to a customer context and / or relative to the accounting data from steps a);
[0053] - accounting data debiting a product account; - accounting data including one or more keywords, preferably dependent on a customer context;
[0054] - accounting data entered after an account closure and / or having a brief description relating to a client context and / or relating to all accounting data from steps a);
[0055] - accounting data whose entry is far from an accounting date relative to a customer context and / or relative to the accounting data resulting from steps a);
[0056] - accounting data presenting a statistically significant impact on an income statement, in particular the accounting data being entered relatively shortly before an account closure;
[0057] - accounting data presenting a statistically significant impact on a revenue account and having been entered after an account closure;
[0058] - accounting data containing round or recurring figures relating to a customer context and / or relating to the accounting data resulting from steps a);
[0059] - accounting data entered several times in relation to a customer context and / or in relation to the accounting data from steps a)
[0060] - accounting data not complying with Benford's law when it is applied to the accounting data from step a);
[0061] - accounting data entered on a non-working day or public holiday;
[0062] - accounting data relating to a current account associated with a creditor;
[0063] - accounting data relating to a suspense account settled at account closure;
[0064] - accounting data relating to a purchase or external charge not paid directly by a bank or fund;
[0065] - accounting data relating to an invoice not recorded as exceptional income or to a credit note not recorded as exceptional expense;
[0066] - verification of the consistency of an accounting scheme over a sales cycle using accounting data from step a);
[0067] - verification of the consistency of an accounting scheme over a purchasing cycle using accounting data from step a);
[0068] - verification of the consistency of an unusual accounting scheme over a cash flow cycle using accounting data from step a), in particular in relation to fixed assets, and / or stocks, and / or equity, and / or VAT (Value Added Tax);
[0069] - verification of the counterbalancing of depreciation accounts, in particular by an advance payment or a recovery;
[0070] - verification of the counterbalancing of provision accounts, in particular by an advance payment or a reversal
[0071] - verification of the counterbalancing of stock accounts, in particular by a movement linked to stocks or provisions;
[0072] - verification of the counterbalancing of tangible fixed asset accounts, in particular by commissioning, acquisitions, sales or scrapping;
[0073] - verification of the counterbalancing of intangible asset accounts, in particular through commissioning, acquisitions, sales or scrapping.
[0074] Fraud detection
[0075] In step c), the data constituting anomalies detected in step b) are provided as input to a machine learning algorithm trained to detect, from input data consisting of accounting data anomalies, accounting data relating to fraud.
[0076] In a particular embodiment, the trained machine learning algorithm also detects accounting data relating to false positives. In an alternative embodiment, a second machine learning algorithm is trained to detect, from input data consisting of accounting data anomalies, accounting data relating to false positives.
[0077] Preferably, the method according to the invention comprises, after step c), a step d) of verification by an operator of the accounting data detected as anomalies by the machine learning algorithm.
[0078] The accounting data verified by the operator can then be labeled and added to a training database of the machine learning algorithm in order to enrich said training database.
[0079] The machine learning algorithm can advantageously be updated from the enriched training database. These steps advantageously make it possible to enrich the training database of the machine learning algorithm and therefore to refine the accuracy of the algorithm in its fraud detection, and possibly false positives.
[0080] The invention can be implemented by means of a network of computers and remotely distributed servers. In particular, the invention can be implemented by means of software as a service (SAAS).
[0081] Of course, the accounting data is then encrypted when it passes between the different servers and computers on the network.
[0082] The method can then be easily deployed. In addition, the method can process more accounting data from different companies, enriching the training database and thus refining detection.
[0083] Device
[0084] The invention also relates to a device for implementing a detection method according to the invention, the device comprising:
[0085] - means of storing accounting data;
[0086] - processing units for implementing compliance tests, all configurable tests, and the machine learning algorithm, preferably a processing unit for implementing a test, or an algorithm, each test being formed of one or more algorithms;
[0087] - at least one application programming interface.
[0088] In a particular embodiment of a device according to the invention, said device is encapsulated in an automatic deployment platform, such as the open source Kubemetes system.
[0089] Such a device allows easy access and installation of the device.
[0090] Brief description of the drawings
[0091] The invention may be better understood by reading the detailed description which follows, non-limiting examples of its implementation, and by examining the following figures: [Fig 1] Figure 1 schematically illustrates the different stages of a method according to the invention,
[0092] [Fig 2a]; [Fig 2b]; [Fig 2c]; [Fig 2d] Figures 2a, 2b, 2c and 2d are block diagrams illustrating the means of implementing the different steps of a method according to the invention,
[0093] [Fig 3] Figure 3 shows a diagram of a device according to the invention.
[0094] In the remainder of the description, elements that are identical or have identical functions bear the same reference sign. Their description is not repeated with respect to each of the figures, only the main differences between the embodiments being mentioned.
[0095] Detailed description
[0096] Figure 1 illustrates a method of detecting fraud from accounting data 10 according to the invention.
[0097] Accounting data may include accounting entries made up of accounting lines, for example from at least one accounting entry file.
[0098] Accounting data can come from several accounting sources, such as accounting periods, accounting ledgers.
[0099] The method comprises the application of compliance rules to accounting data 10 so as to obtain pre-processed accounting data 12.
[0100] The application of compliance rules is carried out by means of a processing unit 22.
[0101] Compliance rules allow you to check that data corresponds to accounting data, particularly with regard to its composition. For example, compliance rules allow you to check whether accounting data includes mandatory fields to be completed for accounting data, whether the format of the fields is respected, for example, is a date in "date" format, to check whether the accounting data is entirely in the same language and whether it is an accepted language with regard to the regulations to be applied.
[0102] These rules allow the measurement of the conformity of data with the accounting criteria commonly accepted in the accounting and financial field and thus determine the quality of said data. Depending on the result obtained by applying a conformity rule to an accounting data, a pre-processing is applied.
[0103] Accounting data that does not respond positively to a compliance rule, in other words that does not comply with the compliance rule, is said to be erroneous.
[0104] Incorrect accounting data may contain one or more errors. In particular, incorrect accounting data may negatively meet one or more compliance rules.
[0105] Preferably, the pre-processing applied to an accounting data depends on the number of errors contained in said data as well as the compliance rule(s) that said accounting data does not comply with.
[0106] In particular, the set of compliance rules may include at least one blocking compliance rule. When accounting data does not comply with a blocking compliance rule, the application of pre-processing to this data preferably involves the deletion of this accounting data.
[0107] When accounting data is erroneous but complies with all the blocking compliance rules, pre-processing may include the allocation of a confidence index, or a weighting of the accounting data considered, the confidence index or the weighting being able to vary according to the number of errors contained in said accounting data.
[0108] When the accounting data is deemed compliant, i.e. it responds positively to the application of the compliance rules, preferably no pre-processing is applied. Alternatively, pre-processing can be applied, for example by indicating a confidence index or a weighting indicating the compliance of the data.
[0109] Compliance rules can be applied sequentially and / or in parallel to each accounting data item. For example, blocking compliance rules can be applied first. It is no longer necessary to apply other compliance rules if an accounting data item violates a blocking compliance rule.
[0110] In particular, the processing unit 22 may comprise several computer modules implemented in parallel, for example by implementing a distributed calculation method. In the Tissue of step a), pre-processed accounting data 12 are advantageously obtained.
[0111] The accounting data 12 from step a) are then subjected to a set of configurable tests so as to detect, among said accounting data 12 from step a), the data constituting anomalies 14.
[0112] All configurable tests can be implemented using a 24-bit processing unit.
[0113] Parameterizable tests advantageously depend on the context in which the accounting data were acquired. In particular, parameterizable tests may depend on the area of the company from which the accounting data 10 originates, the size of the company, and the location of the company's head office.
[0114] The set of configurable tests aims to detect unusual events with regard to the practice of the company from which the accounting data comes.
[0115] For example, configurable tests can detect whether pre-processed accounting data relates to unusual bank account usage, or whether actions were performed on non-working days or outside of business hours.
[0116] Unlike compliance rules, configurable tests analyze accounting data in relation to a context.
[0117] The context can be determined by adapting threshold values or any other parameters that can be entered by an operator such as working days, opening hours, partner bank accounts, usual subcontractors, and / or by statistical analysis of this pre-processed accounting data 12.
[0118] Preferably, configurable tests can be modified by an operator. For example, an operator can modify the working days to be considered within the company, the account closing dates, a limit amount beyond which any higher amount is considered abnormally high or conversely below which any lower amount is considered abnormally low.
[0119] The set of parameterizable tests may include at least five parameterizable tests, preferably at least ten parameterizable tests.
[0120] A parameterized test can be composed of several algorithms.
[0121] Similar to the application of compliance rules, parameterized tests can be applied sequentially and / or in parallel. After applying the parameterized tests, data constituting anomalies is obtained 14.
[0122] Anomalies are typically composed of false positives, gross errors, acceptable errors, and / or fraud.
[0123] In step c), among the anomalies 14, the accounting data relating to fraud 16 are detected, and possibly the accounting data relating to false positives.
[0124] Fraud detection is done using a machine learning algorithm, supervised or unsupervised, trained to detect accounting data relating to fraud from accounting data constituting anomalies.
[0125] The implementation of the machine learning algorithm can be carried out by means of a processing unit 26.
[0126] Preferably, the machine learning algorithm is trained to also detect accounting data relating to false positives among the accounting data constituting anomalies.
[0127] Alternatively, a second machine learning algorithm, different from the machine learning algorithm trained to detect fraud, is trained to detect accounting data relating to false positives among the accounting data constituting anomalies. This second machine learning algorithm can be implemented by means of a processing unit 26, the machine learning algorithms being able to be implemented on different computer modules.
[0128] Data detected as relating to fraud, and where applicable data detected as relating to a false positive, by the machine learning algorithm(s) may be verified by an operator.
[0129] The verified data 18 are then preferably labeled and added to the training database(s) of the machine learning algorithm(s).
[0130] In particular, verification by the operator may consist of an approval or decline of data detected as relating to fraud or, where appropriate, as relating to a false positive.
[0131] Approval or declination can lead to the automatic labeling of the accounting data to enrich a training database 42. The implementation of this verification step thus allows the self-enrichment of the machine learning algorithm(s), thereby improving the performance of the algorithm(s) over time.
[0132] This also allows algorithms to evolve as fraud methods evolve.
[0133] Figures 2a, 2b, 2c and 2d are block diagrams representing the means for implementing the steps of a method according to the invention.
[0134] The processing units 22, 24; 26; 28 may be the same processing unit. One or more processing units may be composed of one or more computer modules, in particular several processors.
[0135] The accounting data detected as relating to fraud or, where appropriate, to a false positive can be displayed on an application programming interface 30. In general, the application programming interface 30 can allow the operator to modify the configurable tests, to view the preprocessed accounting data, to view the accounting data constituting anomalies. Preferably, the visualizations of preprocessed accounting data, accounting data constituting anomalies, accounting data detected as relating to fraud or, where appropriate, to a false positive are presented in the form of a dashboard.
[0136] The method can advantageously be implemented by means of a device by means of a network of computers and remotely distributed servers, in particular by means of software as a service (SAAS).
[0137] Alternatively, the method can be implemented on a single computer, for example in the form of locally deployed software.
[0138] The advantage of a process implemented using SAAS software is the ease of deployment and scaling.
[0139] The invention also relates to a device 1 making it possible to implement the method described above.
[0140] An example of such a device is shown schematically in Figure 2.
[0141] A fraud detection device according to the invention comprises accounting data storage means 40, processing units 20 for implementing compliance tests, all of the configurable tests, and the automatic learning algorithm, and at least one application programming interface 30.
[0142] In the example shown in Figure 2, the device 1 comprises several databases 40 allowing the storage of accounting data, a network of servers and remote computers 20 for the implementation of compliance tests, of all the configurable tests and of the automatic learning algorithm, and of an operator interface 30.
[0143] The accounting data transits between the databases 40 and the networks of servers and computers while being encrypted. Similarly, when the accounting data transits between the databases 40 and the operator interface 30 and / or between the network 20 of servers and computers and the operator interface 30, the accounting data is encrypted.
[0144] The databases 40 may allow the storage of the training database(s) 43 of the machine learning algorithm(s).
[0145] As it now clearly appears, the invention makes it possible in particular to rebalance the frequency of appearance of fraudulent data; balancing the frequency of fraudulent data in relation to all the data provided as input to a machine learning algorithm is necessary to obtain better precision in detecting fraudulent data.
[0146] Furthermore, the invention makes it possible, thanks to the successive processing applied to the accounting data, to precisely detect anomalies constituting accounting data relating to fraud.
[0147] Of course, the invention is not limited to the embodiments which are described.
[0148] In particular, the method according to the invention is advantageously applicable to any type of structured data, i.e. any standardized and organized data such as expense reports, purchasing data, data relating to the determination of e-reputations, data relating to the determination of carbon footprints, data relating to the determination of CSR ("corporate social responsibility") for example by the analysis of ESG (Environmental, Social and Governance) criteria.
[0149] Analyzing other structured data can also improve fraud detection. For example, e-reputation, defined by the French Data Protection Authority (CNIL) as the online image of a company or individual, can help detect the use of registered bank accounts, those located in tax havens, or transactions with dubious organizations. This information can then add bias when detecting fraud among anomalies.
Claims
Claims 1. A computer-implemented method for detecting fraud from accounting data, the method comprising the following successive steps: a) applying a set of predefined compliance rules to the accounting data so as to identify erroneous accounting data, for each erroneous accounting data identified, pre-processing said erroneous accounting data; b) detecting accounting data constituting anomalies by applying a set of configurable tests to the accounting data from step a); c) detecting fraud by submitting the anomalies detected in step b) to a machine learning algorithm trained to detect, from input data consisting of accounting data anomalies, accounting data relating to fraud.
2. Method according to claim 1, in which erroneous accounting data comprises one or more errors, the pre-processing depending on the error or errors.
3. Method according to any one of the preceding claims, in which the pre-processing in step a) comprises a deletion, an attribution of a confidence index, or a weighting of the accounting data considered.
4. Method according to any one of the preceding claims, in which the accounting data comprises accounting entries formed from accounting lines, for example from at least one accounting entry file.
5. Method according to any one of the preceding claims, wherein in step a), the set of predefined compliance rules comprises more than 5 compliance rules, preferably more than 10 rules, preferably more than 20 rules.
6. Method according to any one of the preceding claims, in which the set of predefined compliance rules comprises at least one, preferably several compliance rules from among the following rules: - verification of the information in mandatory fields to be completed in accounting data; - verification of the format of fields to be filled in accounting data, in particular the format of dates; - verification of the presence of a debit amount or a credit amount; - verification of the absence of a debit amount and a credit amount simultaneously; - currency code check for currency amounts, e.g. is it euro corresponds to the code “EUR” and / or the numeric code “978” of the ISO 4217 standard; - verification of the language according to the regulations of the country to which the accounting data must comply, in particular for accounting data relating to French regulations, verification of the absence of words in English; - at least one rule verifying the conformity of accounting data with a national general accounting plan; - at least one rule verifying accounting data with a chart of accounts of a client from which the accounting data comes; - for accounting data being an accounting entry file, verification of the chronology of the accounting entries and / or verification of the absence of sequence inversions and / or verification of the absence of sequence gaps and / or verification of the balance of the entries, in particular the balance of the entries according to the currencies and / or verification of the absence of currency other than euro with a zero amount in currency and / or verification of the dates according to the period concerned by the accounting entry file and / or verification of a lettering subsequent to the accounting entry,and in particular the verification of a lettering subsequent to the validation of the entry and / or verification that the date of an accounting entry is prior to the validation date of said accounting entry and / or verification that a validation period for an accounting entry is less than or equal to 90 days and / or verification of consistency between the different fields of a line of an accounting entry and / or verification of a framework with the general balance and / or verification that an accounting entry includes at least two accounting lines., 7. Method according to any one of the preceding claims, in which at least one compliance rule is a blocking rule involving the deletion of accounting data during pre-processing if said accounting data does not comply with the blocking rule, preferably less than 50% of the compliance rules are blocking, more preferably less than 25% of the compliance rules. . .
8. A method according to any preceding claim, wherein the compliance rules are applied in parallel or sequentially to the accounting data.
9. Method according to any one of the preceding claims, in which the set of parameterizable tests comprises at least 5 tests, preferably at least 10.
10. Method according to claim 9, the set of parameterizable tests being composed of more than 50 algorithms, preferably more than 100 algorithms, preferably more than 200 algorithms, more preferably more than 300 algorithms.
11. Method according to any one of the preceding claims, in which the set of configurable tests comprises all or part of the following tests: - accounting data involving unusual accounts relative to a client context and / or relative to the accounting data resulting from steps a); - accounting data involving rarely used accounts in relation to a client context and / or in relation to the accounting data from steps a); - presence of abnormally high or low amounts relative to a customer context and / or relative to the accounting data from steps a); - accounting data debiting a product account; - accounting data including one or more keywords, preferably dependent on a customer context; - accounting data entered after an account closure and / or having a brief description relating to a client context and / or relating to all accounting data from steps a); - accounting data whose entry is far from an accounting date relative to a customer context and / or relative to the accounting data resulting from steps a); - accounting data presenting a statistically significant impact on an income statement, in particular the accounting data being entered relatively shortly before an account closure; - accounting data presenting a statistically significant impact on a revenue account and having been entered after an account closure; - accounting data containing round or recurring figures relating to a customer context and / or relating to the accounting data resulting from steps a); - accounting data entered several times in relation to a customer context and / or in relation to the accounting data from steps a) - accounting data not complying with Benford's law when it is applied to the accounting data from step a); - accounting data entered on a non-working day or public holiday; - accounting data relating to a current account associated with a creditor; - accounting data relating to a suspense account settled at account closure; - accounting data relating to a purchase or external charge not paid directly by a bank or fund; - accounting data relating to an invoice not recorded as exceptional income or to a credit note not recorded as exceptional expense; - verification of the consistency of an accounting scheme over a sales cycle using accounting data from step a); - verification of the consistency of an accounting scheme over a purchasing cycle using accounting data from step a); - verification of the consistency of an unusual accounting scheme over a cash flow cycle using accounting data from step a), in particular in relation to fixed assets, and / or stocks, and / or equity, and / or VAT (Value Added Tax); - verification of the counterbalancing of depreciation accounts, in particular by an advance payment or a recovery; - verification of the counterbalancing of provision accounts, in particular by an advance payment or a reversal; - verification of the counterbalancing of stock accounts, in particular by a movement linked to stocks or provisions; - verification of the counterbalancing of tangible fixed asset accounts, in particular by commissioning, acquisitions, sales or scrapping; - verification of the counterbalancing of intangible asset accounts, in particular through commissioning, acquisitions, sales or scrapping.
12. Method according to any one of the preceding claims, in which at least one test of the set of configurable tests is configurable by modification of a threshold value, for example the threshold value being representative of an acceptable deviation with regard to a client context.
13. Method according to any one of the preceding claims, comprising after step c), a step d) of verification by an operator of the accounting data detected as anomalies by the machine learning algorithm.
14. Method according to claim 13, in which the accounting data verified by the operator are labeled and then added to a training database. of the machine learning algorithm in order to enrich said training database.
15. Method according to claim 14, the machine learning algorithm being updated from the enriched training database.
16. Device for implementing a method according to any one of the preceding claims comprising: means for storing accounting data; - processing units for implementing compliance tests, all configurable tests, and the machine learning algorithm, preferably a processing unit for implementing a test, or an algorithm, each test being formed of one or more algorithms; - at least one application programming interface.
17. Device according to claim 16, being encapsulated in an automatic deployment platform.
Citation Information
Patent Citations
Data validation
US20190377624A1
Dynamic self-learning system for automatically creating new rules for detecting organizational fraud
WO2017210519A1
Systems and methods for investigation of financial reporting information
US20050222929A1
Methods and Apparatus for Analysing and / or Pre-Processing Financial Accounting Data
US20100205076A1
Automated transaction accounting processing engine and approach
US20120158558A1
