An artificial intelligence enterprise tax risk prediction method and system

Abstract

The application discloses an artificial intelligence enterprise tax risk prediction method and system, relates to the technical field of tax risk prediction, and comprises the following steps: step one, obtaining the associated transaction data of a target enterprise in a historical reporting period and third-party market transaction data in the same period, and respectively extracting transaction attribute information of a to-be-evaluated transaction unit and a potential comparable transaction unit; through a double-deviation-degree characteristic system and a comparable transaction hierarchical correction mechanism, market price information in all comparable transactions can be fully mined under the condition that the number of comparable samples is limited, the problems of insufficient characteristic dimension and low prediction accuracy caused by using only a single deviation degree in the prior art are avoided, low-comparable transactions are weighted and corrected instead of being directly removed through comparability scoring, the contradiction between evaluation accuracy and sample coverage is effectively balanced, and finally the output result of the tax inspection risk prediction model is closer to the actual inspection logic of tax authorities.

Description

Technical Field

[0001] This invention relates to the field of tax risk prediction technology, specifically to an artificial intelligence-based method and system for predicting corporate tax risks. Background Technology

[0002] In the tax compliance management of transfer pricing for related-party transactions, tax authorities require that the prices of related-party transactions comply with the arm's length principle, that is, be consistent with the transaction prices of unrelated third parties under comparable conditions during the same period. Therefore, how to assess the degree of deviation of related-party transaction prices from market prices and predict the risk of tax audits for enterprises is a core requirement in enterprise tax management.

[0003] In existing technologies, the most commonly used comparable uncontrolled price method directly selects third-party transactions that are highly similar to the target related transaction in terms of goods, quantity, and transaction terms from the market database as comparable transactions. The price of these transactions is used as the market reference price to calculate the deviation. However, this method often faces the problem of a severe lack of comparable samples when dealing with special goods or services, resulting in unreliable assessment results. When there are only a few comparable transactions, the fluctuation of the price of a single transaction can have an excessive impact on the assessment results. Some existing technologies attempt to introduce machine learning models to predict tax audit risks based on features such as financial indicators or historical deviations. However, their input features usually only include a single conventional deviation calculated based on highly comparable transactions. They do not stratify comparable transactions according to their comparability quality and treat all comparable transactions the same. This results in the complete discarding of market price information contained in low-comparability transactions, which further exacerbates the problem of insufficient feature information when the comparable sample is already limited. At the same time, the different contributions of high-quality and low-quality comparable information to risk assessment are not distinguished, ultimately resulting in significant deficiencies in the accuracy and practicality of tax audit risk prediction.

[0004] Therefore, in view of this, the present invention proposes an artificial intelligence-based method and system for predicting corporate tax risks to make up for and improve the shortcomings of existing technologies. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides an artificial intelligence-based method and system for predicting corporate tax risks, thereby resolving the corresponding technical issues raised in the background section.

[0006] To achieve the above objectives, the technical solution adopted by this invention is: an artificial intelligence-based method for predicting corporate tax risks, comprising the following steps:

[0007] Step 1: Obtain related party transaction data and third-party market transaction data of the target company during the historical reporting period, and extract transaction attribute information of the transaction unit to be evaluated and potential comparable transaction units respectively;

[0008] Step 2: Based on transaction attribute information, perform multi-dimensional attribute matching on each transaction unit to be evaluated in all potential comparable transaction units, construct an initial comparable transaction set, and calculate the transaction comparability score of each potential comparable transaction unit;

[0009] Step 3: Based on the preset comparability threshold, divide the initial comparable transaction set into a high comparability subset and a low comparability subset. Calculate the normal deviation between the transaction unit to be evaluated and the weighted average price of the high comparability subset, and the comparability correction deviation between the transaction unit to be evaluated and the equivalent market reference price after the comparability score of the low comparability subset.

[0010] Step 4: Based on the regular deviation, comparability-corrected deviation, and corresponding label data of whether or not a tax audit has been conducted for historical enterprise samples, train and construct a tax audit risk prediction model. Input the regular deviation and comparability-corrected deviation of the target enterprise into the tax audit risk prediction model to output the predicted probability value of the target enterprise being subject to tax audit within a preset time window.

[0011] Preferably, step one specifically includes:

[0012] S101. Obtain all related-party transaction data of the target company during the historical reporting period, perform structured parsing on each related-party transaction data, form the transaction attribute vector of the related-party transaction, and mark the related-party transaction as a transaction unit to be evaluated.

[0013] S102. Obtain third-party transaction data in the same period market. The third-party transaction data refers to the transaction data between enterprises that have no relation to the target enterprise that occurred in the same period. Perform structured parsing on each third-party transaction data to form a transaction attribute vector for that third-party transaction and mark that third-party transaction as a potential comparable transaction unit.

[0014] The transaction attribute vector is represented as follows:

[0015] ;

[0016] in, Let be the attribute vector of the i-th transaction;

[0017] For the transaction date;

[0018] Transaction type;

[0019] Product name;

[0020] For the number of transactions;

[0021] The transaction amount;

[0022] and These are the industry codes for the buyer and seller, respectively.

[0023] and These refer to the regions where the buyer and seller are located, respectively.

[0024] Settlement currency;

[0025] These are the payment terms.

[0026] Preferably, step two specifically includes:

[0027] S201. For each transaction unit to be evaluated, multiple attribute matching dimensions are set, including transaction type matching, product name matching, transaction quantity range matching, industry code matching of both parties to the transaction, and region matching of both parties to the transaction.

[0028] S202. For all potential comparable transaction units, compare them one by one with the current transaction unit to be evaluated in each matching dimension. Potential comparable transaction units that meet all the set matching dimensions are included in the initial comparable transaction set of the transaction unit to be evaluated.

[0029] The initial set of comparable transactions is represented as follows:

[0030] ;

[0031] in, This is the initial set of comparable transactions;

[0032] This represents the attribute vector of the transaction unit to be evaluated.

[0033] Let be the attribute vector of the j-th potential comparable transaction unit;

[0034] P is the set of all potentially comparable transaction units;

[0035] This is a multi-dimensional matching function that returns True when all dimensions meet the matching conditions.

[0036] S203, For the initial set of comparable transactions For each potential comparable transaction unit j, calculate its transaction comparability score with the transaction unit i to be evaluated. The calculation formula is as follows:

[0037] ;

[0038] in, The transaction comparability score is given to the j-th potential comparable transaction unit relative to the transaction unit to be evaluated;

[0039] K is the total number of matching dimensions;

[0040] Let be the preset weight for the k-th matching dimension, and ;

[0041] Let be the similarity function on the k-th matching dimension;

[0042] When the k-th dimension meets the matching condition ,otherwise, ;

[0043] When all K dimensions are satisfied The fewer the dimensions to satisfy, The smaller.

[0044] Preferably, step three specifically includes:

[0045] S301, Set the preset comparability threshold The initial set of comparable transactions Each potential comparable transaction unit is scored according to its transaction comparability. and Compare, Potentially comparable transaction units are classified into a high-comparability subset. ,Will Potentially comparable transaction units are classified into a subset of low comparability. ;

[0046] S302. Calculate the transaction price of the transaction unit to be evaluated. Highly comparable subsets Normal deviation between the weighted average market prices of comparable transactions The calculation formula is as follows:

[0047] ;

[0048] ;

[0049] in, The weighted average market price of the transaction comparability score for the high comparability subset;

[0050] Let be the transaction price of the j-th comparable transaction in the highly comparable subset;

[0051] The corresponding transaction comparability score is used as the weight;

[0052] The standard deviation represents the percentage deviation of the transaction price under evaluation from the weighted average price of highly comparable transactions.

[0053] S303, Based on low-comparability subsets Comparability score of each potential comparable transaction unit The transaction prices of the low comparability subset are adjusted for comparability weighting to obtain the equivalent market reference price. The calculation formula is as follows:

[0054] ;

[0055] S304. Calculate the transaction price of the transaction unit to be evaluated. Equivalent market reference price Comparability Correction Deviation The calculation formula is as follows:

[0056] .

[0057] Preferably, step four specifically includes:

[0058] S401. Collect data from the tax history database for N enterprise samples during their respective historical reporting periods. For each enterprise sample n, extract the regular deviation of each transaction unit to be evaluated during that reporting period. and comparability correction deviation And obtain the label of whether the enterprise sample was investigated by the tax authorities for transfer pricing issues within a preset time window T after the reporting period. ,in, This indicates that the entity is under investigation. This indicates that the audit was not conducted.

[0059] S402, with standard deviation and comparability correction deviation Using the input feature vector and the label y as the output target, a tax audit risk prediction model is trained using the logistic regression algorithm. The formula is as follows:

[0060] ;

[0061] in, For the Sigmoid function;

[0062] For the intercept term;

[0063] and These are the regression coefficients;

[0064] Given the normal deviation and the comparability-corrected deviation, this is the predicted probability of the company being subject to a tax audit.

[0065] S403. Calculate the regular deviation of the target company during the current reporting period. and comparability correction deviation Substituting the values ​​into the trained tax audit risk prediction model, the predicted probability of the target company being subject to a tax audit due to transfer pricing within a preset time window T is calculated and output. The calculation formula is as follows:

[0066] ;

[0067] in, The range of values ​​is , The closer the value is to 1, the higher the risk that the target company will be subject to tax audit within the preset time window.

[0068] Preferably, the process of training and constructing the tax audit risk prediction model using the logistic regression algorithm further includes:

[0069] The collected sample data from N enterprises are divided into training set and test set in an 8:2 ratio. The training set data is used for parameter fitting, and the test set data is used to evaluate the model's AUC index. When the AUC is greater than the preset threshold, the model is considered to have completed training and is effective.

[0070] An artificial intelligence-based enterprise tax risk prediction system includes:

[0071] The information extraction module is used to obtain related transaction data of the target company during the historical reporting period and third-party transaction data of the market during the same period, and to extract transaction attribute information of the transaction unit to be evaluated and the potential comparable transaction units respectively;

[0072] The scoring calculation module, based on transaction attribute information, performs multi-dimensional attribute matching on each transaction unit to be evaluated in all potential comparable transaction units, constructs an initial set of comparable transactions, and calculates the transaction comparability score of each potential comparable transaction unit.

[0073] The dual deviation calculation module divides the initial comparable transaction set into a high-comparability subset and a low-comparability subset based on a preset comparability threshold. It calculates the regular deviation between the transaction unit to be evaluated and the weighted average price of the high-comparability subset, as well as the comparability-corrected deviation between the transaction unit to be evaluated and the equivalent market reference price after the comparability score of the low-comparability subset.

[0074] The prediction output module trains and constructs a tax audit risk prediction model based on the regular deviation, comparability-corrected deviation, and corresponding label data of whether or not a company is subject to tax audit based on historical enterprise samples. The regular deviation and comparability-corrected deviation of the target company are input into the tax audit risk prediction model to output the predicted probability value of the target company being subject to tax audit within a preset time window.

[0075] Compared with existing technologies, the beneficial effects of this invention are as follows: By acquiring related-party transaction data and third-party market transaction data of the target enterprise during the historical reporting period, transaction attribute information of the transaction unit to be evaluated and potential comparable transaction units is extracted respectively. Based on the transaction attribute information, multi-dimensional attribute matching is performed on each transaction unit to be evaluated among all potential comparable transaction units to construct an initial comparable transaction set and calculate the transaction comparability score of each potential comparable transaction unit. Based on a preset comparability threshold, the initial comparable transaction set is divided into a high-comparability subset and a low-comparability subset. The conventional deviation between the transaction unit to be evaluated and the weighted average price of the high-comparability subset, and the comparability correction deviation between the transaction unit to be evaluated and the equivalent market reference price after correction by the transaction comparability score of the low-comparability subset are calculated respectively. Based on the conventional deviation, comparability correction deviation, and corresponding whether the transaction unit to be evaluated is evaluated and the ... The model uses labeled data of companies subject to tax audits to train and construct a tax audit risk prediction model. The model inputs the target company's regular deviation and comparability-corrected deviation into the model to output a predicted probability of the target company being audited within a preset time window. Through a dual-deviation feature system and a tiered correction mechanism for comparable transactions, it can fully extract market price information from all comparable transactions even with a limited number of comparable samples. This avoids the problems of insufficient feature dimensions and low prediction accuracy caused by using only a single deviation in existing technologies. Furthermore, by using comparability scoring to weight and correct low-comparability transactions instead of directly eliminating them, it effectively balances the contradiction between assessment accuracy and sample coverage. Ultimately, the output of the tax audit risk prediction model is closer to the actual audit logic of tax authorities, providing companies with more accurate and practically valuable tax risk warnings. Attached Figure Description

[0076] Figure 1 This is a schematic diagram of the overall structure of a preferred embodiment of the present invention;

[0077] Figure 2 This is a schematic diagram of the system structure shown in this invention. Detailed Implementation

[0078] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0079] Embodiments of the present invention:

[0080] Please refer to Figures 1 to 2 As shown, an artificial intelligence-based method and system for predicting corporate tax risks includes the following steps:

[0081] Step 1: Obtain related party transaction data and third-party market transaction data of the target company during the historical reporting period, and extract transaction attribute information of the transaction unit to be evaluated and potential comparable transaction units respectively;

[0082] Step 2: Based on transaction attribute information, perform multi-dimensional attribute matching on each transaction unit to be evaluated in all potential comparable transaction units, construct an initial comparable transaction set, and calculate the transaction comparability score of each potential comparable transaction unit;

[0083] Step 3: Based on the preset comparability threshold, divide the initial comparable transaction set into a high comparability subset and a low comparability subset. Calculate the normal deviation between the transaction unit to be evaluated and the weighted average price of the high comparability subset, and the comparability correction deviation between the transaction unit to be evaluated and the equivalent market reference price after the comparability score of the low comparability subset.

[0084] Step 4: Based on the regular deviation, comparability-corrected deviation, and corresponding label data of whether or not a tax audit has been conducted for historical enterprise samples, train and construct a tax audit risk prediction model. Input the regular deviation and comparability-corrected deviation of the target enterprise into the tax audit risk prediction model to output the predicted probability value of the target enterprise being subject to tax audit within a preset time window.

[0085] Step one specifically includes:

[0086] S101. Obtain all related-party transaction data of the target company during the historical reporting period, perform structured parsing on each related-party transaction data, form the transaction attribute vector of the related-party transaction, and mark the related-party transaction as a transaction unit to be evaluated.

[0087] S102. Obtain third-party transaction data in the market during the same period. The third-party transaction data refers to the transaction data between companies that are not related to the target company during the same period. Perform structured parsing on each third-party transaction data to form the transaction attribute vector of the third-party transaction and mark the third-party transaction as a potential comparable transaction unit.

[0088] The transaction attribute vector is represented as follows:

[0089] ;

[0090] in, Let be the attribute vector of the i-th transaction;

[0091] For the transaction date;

[0092] Transaction type;

[0093] Product name;

[0094] For the number of transactions;

[0095] The transaction amount;

[0096] and These are the industry codes for the buyer and seller, respectively.

[0097] and These refer to the regions where the buyer and seller are located, respectively.

[0098] Settlement currency;

[0099] These are the payment terms.

[0100] Step two specifically includes:

[0101] S201. For each transaction unit to be evaluated, set multiple attribute matching dimensions, including transaction type matching, product name matching, transaction quantity range matching, industry code matching of both parties, and region matching of both parties.

[0102] S202. For all potential comparable transaction units, compare them one by one with the current transaction unit to be evaluated in each matching dimension. Potential comparable transaction units that meet all the set matching dimensions are included in the initial comparable transaction set of the transaction unit to be evaluated.

[0103] The initial set of comparable transactions is represented as follows:

[0104] ;

[0105] in, This is the initial set of comparable transactions;

[0106] This represents the attribute vector of the transaction unit to be evaluated.

[0107] Let be the attribute vector of the j-th potential comparable transaction unit;

[0108] P is the set of all potentially comparable transaction units;

[0109] This is a multi-dimensional matching function that returns True when all dimensions meet the matching conditions.

[0110] S203, For the initial set of comparable transactions For each potential comparable transaction unit j, calculate its transaction comparability score with the transaction unit i to be evaluated. The calculation formula is as follows:

[0111] ;

[0112] in, The transaction comparability score is given to the j-th potential comparable transaction unit relative to the transaction unit to be evaluated;

[0113] K is the total number of matching dimensions;

[0114] Let be the preset weight for the k-th matching dimension, and ;

[0115] Let be the similarity function on the k-th matching dimension;

[0116] When the k-th dimension meets the matching condition ,otherwise, ;

[0117] When all K dimensions are satisfied The fewer the dimensions to satisfy, The smaller.

[0118] Step three specifically includes:

[0119] S301, Set the preset comparability threshold The initial set of comparable transactions Each potential comparable transaction unit is scored according to its transaction comparability. and Compare, Potentially comparable transaction units are classified into a high-comparability subset. ,Will Potentially comparable transaction units are classified into a subset of low comparability. ;

[0120] S302. Calculate the transaction price of the transaction unit to be evaluated. Highly comparable subsets Normal deviation between the weighted average market prices of comparable transactions The calculation formula is as follows:

[0121] ;

[0122] ;

[0123] in, The weighted average market price of the transaction comparability score for the high comparability subset;

[0124] Let be the transaction price of the j-th comparable transaction in the highly comparable subset;

[0125] The corresponding transaction comparability score is used as the weight;

[0126] The standard deviation represents the percentage deviation of the transaction price under evaluation from the weighted average price of highly comparable transactions.

[0127] S303, Based on low-comparability subsets Comparability score of each potential comparable transaction unit The transaction prices of the low comparability subset are adjusted for comparability weighting to obtain the equivalent market reference price. The calculation formula is as follows:

[0128] ;

[0129] S304. Calculate the transaction price of the transaction unit to be evaluated. Equivalent market reference price Comparability Correction Deviation The calculation formula is as follows:

[0130] .

[0131] Step four specifically includes:

[0132] S401. Collect data from the tax history database for N enterprise samples during their respective historical reporting periods. For each enterprise sample n, extract the regular deviation of each transaction unit to be evaluated during that reporting period. and comparability correction deviation And obtain the label of whether the enterprise sample was investigated by the tax authorities for transfer pricing issues within a preset time window T after the reporting period. ,in, This indicates that the entity is under investigation. This indicates that the audit was not conducted.

[0133] S402, with standard deviation and comparability correction deviation Using the input feature vector and the label y as the output target, a tax audit risk prediction model is trained using the logistic regression algorithm. The formula is as follows:

[0134] ;

[0135] in, For the Sigmoid function;

[0136] For the intercept term;

[0137] and These are the regression coefficients;

[0138] Given the normal deviation and the comparability-corrected deviation, this is the predicted probability of the company being subject to a tax audit.

[0139] S403. Calculate the regular deviation of the target company during the current reporting period. and comparability correction deviation Substituting the values ​​into the trained tax audit risk prediction model, the predicted probability of the target company being subject to a tax audit due to transfer pricing within a preset time window T is calculated and output. The calculation formula is as follows:

[0140] ;

[0141] in, The range of values ​​is , The closer the value is to 1, the higher the risk that the target company will be subject to tax audit within the preset time window.

[0142] The process of training and constructing a tax audit risk prediction model using the logistic regression algorithm also includes:

[0143] The collected sample data from N enterprises are divided into training set and test set in an 8:2 ratio. The training set data is used for parameter fitting, and the test set data is used to evaluate the model's AUC index. When the AUC is greater than the preset threshold, the model is considered to have completed training and is effective.

[0144] An artificial intelligence-based enterprise tax risk prediction system includes:

[0145] The information extraction module is used to obtain related transaction data of the target company during the historical reporting period and third-party transaction data of the market during the same period, and to extract transaction attribute information of the transaction unit to be evaluated and the potential comparable transaction units respectively;

[0146] The scoring calculation module, based on transaction attribute information, performs multi-dimensional attribute matching on each transaction unit to be evaluated in all potential comparable transaction units, constructs an initial set of comparable transactions, and calculates the transaction comparability score of each potential comparable transaction unit.

[0147] The dual deviation calculation module divides the initial comparable transaction set into a high-comparability subset and a low-comparability subset based on a preset comparability threshold. It calculates the regular deviation between the transaction unit to be evaluated and the weighted average price of the high-comparability subset, as well as the comparability-corrected deviation between the transaction unit to be evaluated and the equivalent market reference price after the comparability score of the low-comparability subset.

[0148] The prediction output module trains and constructs a tax audit risk prediction model based on the regular deviation, comparability-corrected deviation, and corresponding label data of whether or not a company is subject to tax audit based on historical enterprise samples. The regular deviation and comparability-corrected deviation of the target company are input into the tax audit risk prediction model to output the predicted probability value of the target company being subject to tax audit within a preset time window.

[0149] By acquiring related-party transaction data and third-party market transaction data of the target company during the historical reporting period, transaction attribute information of the transaction unit to be evaluated and potential comparable transaction units is extracted. Based on the transaction attribute information, multi-dimensional attribute matching is performed on each transaction unit to be evaluated among all potential comparable transaction units to construct an initial comparable transaction set and calculate the transaction comparability score of each potential comparable transaction unit. Based on a preset comparability threshold, the initial comparable transaction set is divided into a high-comparability subset and a low-comparability subset. The conventional deviation between the transaction unit to be evaluated and the weighted average price of the high-comparability subset, and the comparability correction deviation between the transaction unit to be evaluated and the equivalent market reference price after correction by the transaction comparability score of the low-comparability subset are calculated. Based on the conventional deviation, comparability correction deviation, and the corresponding number of tags indicating whether or not the company has been subject to tax audit, the transaction is further analyzed. According to reports, a tax audit risk prediction model has been developed. The model inputs the target company's regular deviation and comparability-corrected deviation into the model to output a predicted probability of the target company being audited within a preset time window. Through a dual-deviation feature system and a tiered correction mechanism for comparable transactions, it can fully extract market price information from all comparable transactions even with a limited number of comparable samples. This avoids the problems of insufficient feature dimensions and low prediction accuracy caused by using only a single deviation in existing technologies. Furthermore, by using comparability scoring to weight and correct low-comparability transactions instead of directly eliminating them, it effectively balances the contradiction between assessment accuracy and sample coverage. Ultimately, the output of the tax audit risk prediction model is closer to the actual audit logic of tax authorities, providing companies with more accurate and practically valuable tax risk warnings.

[0150] The size of the interval and threshold is set to facilitate comparison. The size of the threshold depends on the amount of sample data and the number of bases set by those skilled in the art for each set of sample data; as long as it does not affect the ratio between the parameter and the quantized value.

[0151] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0152] In the two embodiments provided in this application, it should be understood that the disclosed apparatus and system can be implemented in other ways; for example, the apparatus embodiments described above are merely illustrative, and the division of modules is merely a logical functional division. In actual implementation, there may be other division methods, such as multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed; furthermore, the coupling or direct coupling or communication connection between the shown or discussed mutuals can be through some interfaces, and the indirect coupling or communication connection between the apparatus or modules can be electrical, mechanical or other forms.

[0153] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting corporate tax risks using artificial intelligence, characterized in that, Includes the following steps: Step 1: Obtain related party transaction data and third-party market transaction data of the target company during the historical reporting period, and extract transaction attribute information of the transaction unit to be evaluated and potential comparable transaction units respectively; Step 2: Based on transaction attribute information, perform multi-dimensional attribute matching on each transaction unit to be evaluated in all potential comparable transaction units, construct an initial comparable transaction set, and calculate the transaction comparability score of each potential comparable transaction unit; Step 3: Based on the preset comparability threshold, divide the initial comparable transaction set into a high comparability subset and a low comparability subset. Calculate the normal deviation between the transaction unit to be evaluated and the weighted average price of the high comparability subset, and the comparability correction deviation between the transaction unit to be evaluated and the equivalent market reference price after the comparability score of the low comparability subset. Step 4: Based on the regular deviation, comparability-corrected deviation, and corresponding label data of whether or not a tax audit has been conducted for historical enterprise samples, train and construct a tax audit risk prediction model. Input the regular deviation and comparability-corrected deviation of the target enterprise into the tax audit risk prediction model to output the predicted probability value of the target enterprise being subject to tax audit within a preset time window.

2. The artificial intelligence-based enterprise tax risk prediction method according to claim 1, characterized in that, Step one specifically includes: S101. Obtain all related-party transaction data of the target company during the historical reporting period, perform structured parsing on each related-party transaction data, form the transaction attribute vector of the related-party transaction, and mark the related-party transaction as a transaction unit to be evaluated. S102. Obtain third-party transaction data in the same period market. The third-party transaction data refers to the transaction data between enterprises that have no relation to the target enterprise that occurred in the same period. Perform structured parsing on each third-party transaction data to form a transaction attribute vector for that third-party transaction and mark that third-party transaction as a potential comparable transaction unit. The transaction attribute vector is represented as follows: ； in, Let be the attribute vector of the i-th transaction; For the transaction date; Transaction type; Product name; For the number of transactions; The transaction amount; and These are the industry codes for the buyer and seller, respectively. and These refer to the regions where the buyer and seller are located, respectively. Settlement currency; These are the payment terms.

3. The enterprise tax risk prediction method based on artificial intelligence according to claim 1, characterized in that, Step two specifically includes: S201. For each transaction unit to be evaluated, multiple attribute matching dimensions are set, including transaction type matching, product name matching, transaction quantity range matching, industry code matching of both parties to the transaction, and region matching of both parties to the transaction. S202. For all potential comparable transaction units, compare them one by one with the current transaction unit to be evaluated in each matching dimension. Potential comparable transaction units that meet all the set matching dimensions are included in the initial comparable transaction set of the transaction unit to be evaluated. The initial set of comparable transactions is represented as follows: ； in, This is the initial set of comparable transactions; This represents the attribute vector of the transaction unit to be evaluated. Let be the attribute vector of the j-th potential comparable transaction unit; P is the set of all potentially comparable transaction units; This is a multi-dimensional matching function that returns True when all dimensions meet the matching conditions. S203, For the initial set of comparable transactions For each potential comparable transaction unit j, calculate its transaction comparability score with the transaction unit i to be evaluated. The calculation formula is as follows: ； in, The transaction comparability score is given to the j-th potential comparable transaction unit relative to the transaction unit to be evaluated; K is the total number of matching dimensions; Let be the preset weight for the k-th matching dimension, and ; Let be the similarity function on the k-th matching dimension; When the k-th dimension meets the matching condition ,otherwise, ; When all K dimensions are satisfied The fewer the dimensions to satisfy, The smaller.

4. The enterprise tax risk prediction method based on artificial intelligence according to claim 1, characterized in that, Step three specifically includes: S301, Set the preset comparability threshold The initial set of comparable transactions Each potential comparable transaction unit is scored according to its transaction comparability. and Compare, Potentially comparable transaction units are classified into a high comparability subset. ,Will Potentially comparable transaction units are classified into a subset of low comparability. ; S302. Calculate the transaction price of the transaction unit to be evaluated. Highly comparable subsets Normal deviation between the weighted average market prices of comparable transactions The calculation formula is as follows: ； ； in, The weighted average market price of the transaction comparability score for the high comparability subset; Let be the transaction price of the j-th comparable transaction in the highly comparable subset; The corresponding transaction comparability score is used as the weight; The standard deviation represents the percentage deviation of the transaction price under evaluation from the weighted average price of highly comparable transactions. S303, Based on low-comparability subsets Comparability score of each potential comparable transaction unit The transaction prices of the low comparability subset are adjusted for comparability weighting to obtain the equivalent market reference price. The calculation formula is as follows: ； S304. Calculate the transaction price of the transaction unit to be evaluated. Equivalent market reference price Comparability Correction Deviation The calculation formula is as follows: 。 5. The enterprise tax risk prediction method based on artificial intelligence according to claim 1, characterized in that, Step four specifically includes: S401. Collect data from the tax history database for N enterprise samples during their respective historical reporting periods. For each enterprise sample n, extract the regular deviation of each transaction unit to be evaluated during that reporting period. and comparability correction deviation And obtain the label of whether the enterprise sample was investigated by the tax authorities for transfer pricing issues within a preset time window T after the reporting period. ,in, This indicates that the entity is under investigation. This indicates that the audit was not conducted. S402, with standard deviation and comparability correction deviation Using the input feature vector and the label y as the output target, a tax audit risk prediction model is trained using the logistic regression algorithm. The formula is as follows: ； in, For the Sigmoid function; For the intercept term; and These are the regression coefficients; Given the normal deviation and the comparability-corrected deviation, this is the predicted probability of the company being subject to a tax audit. S403. Calculate the regular deviation of the target company during the current reporting period. and comparability correction deviation Substituting the values ​​into the trained tax audit risk prediction model, the predicted probability of the target company being subject to a tax audit due to transfer pricing within a preset time window T is calculated and output. The calculation formula is as follows: ； in, The range of values ​​is , The closer the value is to 1, the higher the risk that the target company will be subject to tax audit within the preset time window.

6. The artificial intelligence-based enterprise tax risk prediction method according to claim 5, characterized in that, The process of training and constructing a tax audit risk prediction model using the logistic regression algorithm also includes: The collected sample data from N enterprises are divided into training set and test set in an 8:2 ratio. The training set data is used for parameter fitting, and the test set data is used to evaluate the model's AUC index. When the AUC is greater than the preset threshold, the model is considered to have completed training and is effective.

7. An artificial intelligence-based enterprise tax risk prediction system, applied to the artificial intelligence-based enterprise tax risk prediction method according to any one of claims 1-6, characterized in that, include: The information extraction module is used to obtain related transaction data of the target company during the historical reporting period and third-party transaction data of the market during the same period, and to extract transaction attribute information of the transaction unit to be evaluated and the potential comparable transaction units respectively; The scoring calculation module, based on transaction attribute information, performs multi-dimensional attribute matching on each transaction unit to be evaluated in all potential comparable transaction units, constructs an initial set of comparable transactions, and calculates the transaction comparability score of each potential comparable transaction unit. The dual deviation calculation module divides the initial comparable transaction set into a high-comparability subset and a low-comparability subset based on a preset comparability threshold. It calculates the regular deviation between the transaction unit to be evaluated and the weighted average price of the high-comparability subset, as well as the comparability-corrected deviation between the transaction unit and the equivalent market reference price after the comparability score of the low-comparability subset. The prediction output module trains and constructs a tax audit risk prediction model based on the regular deviation, comparability-corrected deviation, and corresponding label data of whether or not a company is subject to tax audit based on historical enterprise samples. The regular deviation and comparability-corrected deviation of the target company are input into the tax audit risk prediction model to output the predicted probability value of the target company being subject to tax audit within a preset time window.

Images