Financial data causal relationship mining method and system for intelligent auditing

Abstract

The invention discloses a financial data causal relationship mining method and system for intelligent auditing. The method comprises the steps of obtaining and preprocessing structured financial data and unstructured text data; fusing domain knowledge and a data-driven algorithm to construct a financial causal graph; using a large language model to identify and quantify management layer intervention variables; constructing a structured causal model and executing anti-fact deduction to obtain a financial index prediction value under a non-intervention scene; and comparing the actual observed value with the predicted value, quantifying the manipulation effect and generating a risk clue. The invention further discloses a corresponding system which comprises a data acquisition and preprocessing function module, a financial causal diagram construction function module, a management layer intervention variable recognition and quantification function module, an anti-fact deduction function module, a manipulation effect mining function module and the like. According to the method, analysis from abnormality discovery to causal attribution can be realized, and the net influence of management layer decisions on financial results is quantified, so that the recognition capability and auditing efficiency of complex profit manipulation behaviors are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of financial data processing and artificial intelligence auditing, and particularly relates to a financial data causal relationship mining method and system for intelligent auditing. BACKGROUND

[0002] In the auditing of enterprise financial statements, one of the core tasks for auditors is to identify and assess the risk of financial misstatements, especially those caused by management fraud. Existing auditing methods, such as statistical sampling, trend analysis, financial ratio analysis, and internal control testing, mainly rely on the analysis of the correlation between financial data. Auditors identify potential risk points by finding abnormal fluctuations or inconsistent relationships between financial indicators. For example, when it is found that the company's sales revenue has increased significantly while its accounts receivable turnover rate has decreased significantly, the abnormal correlation will guide auditors to perform further audit procedures.

[0003] However, with the increasing complexity of business operations and the diversification of financial structures, especially the concealment and compliance of profit manipulation methods, the traditional auditing method relying only on correlation analysis has shown its limitations in practice. Some advanced profit manipulation behaviors are not implemented through direct violation of accounting standards, but through careful selection of accounting policies, use of accounting estimates, and design of complex transaction structures within the framework allowed by the rules. In such cases, various financial indicators may still maintain a seemingly reasonable correlation on the surface, and traditional auditing methods are difficult to penetrate the appearance of compliance and discover the real economic motivation and manipulation intention hidden behind. For example, management may use complex related party transactions or set up special purpose entities to recognize revenue or transfer debt in advance, which may be in line with the requirements of accounting standards in form, but in essence distorts the true financial position and operating results of the enterprise. Traditional audit procedures are difficult to attribute a specific financial result, such as abnormal growth of profits, to a specific accounting decision or transaction arrangement of the management, and the audit judgment often stays at the level of possible risk, lacking direct and causally directed evidence support, which constitutes a bottleneck for the improvement of audit efficiency and quality. Similarly, in the fields of enterprise internal control, financial investment decision-making, and financial supervision, there are also challenges in accurately assessing the true operating status and potential risks only through the appearance of financial data. For example, it is difficult for internal audit departments to effectively prevent and discover management compliance fraud; it is difficult for investment institutions to remove the modification effect of accounting policy selection on profits when performing valuation; and it is also difficult for regulatory agencies to efficiently identify complex profit manipulation clues from massive data when performing large-scale financial statement screening.

[0004] In recent years, artificial intelligence technology has begun to be applied in the field of auditing in order to improve the automation and intelligence level of auditing work. At present, most intelligent auditing tools mainly focus on using machine learning algorithms for anomaly detection. These tools can efficiently process massive transaction data and identify abnormal points that do not conform to historical patterns or expected behavior. Although this improves the efficiency of finding abnormal clues to some extent, the underlying logic is still based on correlation analysis or pattern matching, which fails to fundamentally solve the attribution problem. That is, existing technologies can indicate the location of abnormalities, but it is difficult to explain the causes of abnormalities, and it is even more difficult to quantify the specific impact of a particular behavior on financial results. Therefore, there is a need in the art for a new technical solution that can go beyond the traditional "correlation" analysis paradigm and establish a "causality" reasoning path from management accounting decisions to the final financial results, thereby enabling effective identification and quantification of deep, compliant profit manipulation behaviors. SUMMARY

[0005] The technical problem to be solved by the present application is that the existing auditing methods rely on correlation analysis and are difficult to effectively identify deep, compliant profit manipulation behaviors. Therefore, the present application provides a financial data causality mining system and method for intelligent auditing that can penetrate the compliant appearance and reveal the causal chain behind profit manipulation.

[0006] To solve the above technical problems, one aspect of the present application provides a financial data causality mining method for intelligent auditing, which comprises the following steps: Step one, multi-source heterogeneous financial data acquisition and preprocessing. This step establishes a standardized data basis for subsequent causality mining. First, the system acquires structured and unstructured data related to the target auditing entity from multiple data sources. Specifically, structured data includes but is not limited to various subject data in enterprise historical annual and quarterly financial statements such as balance sheets, income statements, and cash flow statements, as well as detailed transaction flow data recorded in enterprise resource planning systems. In addition, macroeconomic indicators and industry average data can be acquired as external environmental variables.

[0007] Unstructured data mainly includes the management discussion and analysis section in the annual report, various temporary announcements, news releases, industry research reports, and third-party financial news reports. After data acquisition, the system performs a series of preprocessing operations. For structured data, data cleaning is performed, including using specific algorithms to handle missing values and outliers, such as using cubic spline interpolation to fill in missing quarterly data in financial statements, and using unsupervised learning algorithms such as isolation forest to identify and correct extreme abnormal financial ratio data. After that, all financial subject data are standardized and aligned to form a unified time series data set.

[0008] For unstructured data, natural language processing is performed, specifically including sentence segmentation, word segmentation, stop word removal, part-of-speech tagging, and named entity recognition. In particular, the system employs a pre-trained named entity recognition model in the financial field to accurately identify and extract specific financial subjects (such as "goodwill", "deferred tax assets"), accounting policies (such as "income recognition point"), key figures (such as "chief financial officer"), and specific economic events (such as "asset restructuring", "stock incentive") and other entity information mentioned in the text. Finally, all preprocessed structured and unstructured data is integrated and stored in a unified data warehouse.

[0009] Step two, perform domain knowledge and data-driven fusion-based financial causal diagram construction. This step constructs a financial causal diagram that can describe the causal relationships between various financial indicators, business activities, and external factors of an enterprise. The financial causal diagram is a directed acyclic graph (DAG) in form, where the nodes represent various variables, and the directed edges between the nodes represent the direct causal relationships between the variables. The present invention constructs a financial causal diagram by fusing domain knowledge and data-driven causal discovery algorithms, providing a foundation for subsequent causal inference.

[0010] The specific construction process is as follows: First, based on recognized accounting standards, accounting identities, and macroeconomic theories, an initial domain knowledge causal diagram is constructed. The initial diagram contains the relatively stable financial causal relationships recognized in the field, for example, according to business logic, it can be determined that "operating income" is the direct cause of "operating profit", "sales expenses" is the direct cause of "operating profit", and "cash paid for the purchase of fixed assets, intangible assets, and other long-term assets" is one of the direct causes of the change in the book value of "fixed assets". The initial graph constructed based on domain knowledge provides an initial structure and strong constraints for the construction of the entire causal diagram, effectively avoiding the pseudo-causal relationships that may be learned by purely data-driven algorithms. Subsequently, based on the domain knowledge causal diagram, a data-driven causal discovery algorithm is used to learn the structured time series data set prepared in step one to discover hidden, non-obvious potential causal relationships in the data, and to supplement and modify the initial graph. Preferably, a constraint-based causal discovery algorithm such as the PC algorithm (Peter-Clark Algorithm) or the FCI algorithm (Fast Causal Inference Algorithm) can be used. This type of algorithm infers the causal structure between variables by performing a series of conditional independence tests on the data, and has the advantages of strong processing capability for high-dimensional data and fewer assumptions about data distribution, making it suitable for complex financial time series data.

[0011] For example, by analyzing historical data, the algorithm might discover that for a company in a specific industry, changes in "inventory turnover" significantly affect the ratio of "selling expenses" to revenue in the next accounting period. After passing a statistical significance test, this causal relationship edge can be added to the initial knowledge causal graph. Ultimately, the system integrates the data-driven causal relationships with the initial knowledge causal graph to form a more complete and accurate financial causal graph.

[0012] Step three involves identifying and quantifying management intervention behaviors based on a large-scale language model. The core of this step is to automatically identify, from the unstructured text data processed in Step one, accounting decisions or operational behaviors led by management that may have a substantial impact on the company's financial results, and to identify and quantify these behaviors as "intervention variables." This invention, by employing a large-scale language model fine-tuned in the financial domain, can automatically identify and quantify management intervention variables from unstructured text.

[0013] Preferably, this step employs a large-scale language model that has undergone instruction fine-tuning on a massive corpus of professional financial terminology, such as a language model based on the Transformer architecture. To achieve accurate identification of intervention behaviors, the system pre-constructs a specialized instruction fine-tuning dataset, which contains a large number of fragments extracted from texts such as "management discussions and analyses," and sentences describing manipulable accounting decisions or non-routine business practices have been annotated by senior audit experts.

[0014] For example, sentences describing changes in revenue recognition policies, adjustments to asset impairment provision methods, significant non-routine related-party transactions, or changes in the boundary between capitalization and expensing of R&D expenditures are labeled as positive samples. When performing the recognition task, pre-processed unstructured text is input into this fine-tuned large-scale language model. The model is trained to perform sequence labeling or text classification tasks to identify and extract two types of key information: first, the type of intervention behavior, such as identifying specific behaviors like "accounts receivable factoring," "sale-leaseback transactions," "goodwill impairment testing," or "R&D expenditure capitalization"; second, the relevant parameters of the intervention behavior, such as the specific amount involved, the time of the transaction, and the counterparty.

[0015] For example, when the model processes the sentence "During this reporting period, the company changed some eligible R&D investments from expense treatment to capitalization treatment, which increased intangible assets by approximately RMB 50 million in the current period," it can automatically identify this as an intervention of type "R&D expenditure capitalization" and extract "RMB 50 million" as the quantitative parameter of this intervention. Finally, the system precisely maps each identified intervention variable to one or more corresponding nodes in the financial causal graph constructed in step two, serving as the intervention point for subsequent counterfactual inference.

[0016] Step four involves performing counterfactual inference based on a structured causal model. This step is used to achieve causal attribution by simulating financial outcomes under no-intervention conditions through counterfactual reasoning. This invention transforms financial causal diagrams into structured causal models and utilizes counterfactual inference to quantify the true impact of management intervention on key financial indicators, thus achieving a shift from correlation analysis to causal attribution.

[0017] First, the financial causal graph constructed in step two is transformed into a structured causal model (SCM). In an SCM, the value of each variable is determined by a function whose inputs are the values ​​of all its direct causes (i.e., parent nodes) in the causal graph, and an exogenous random disturbance term representing unobserved factors. These causal functions can be learned in various ways. Preferably, machine learning models with nonlinear modeling capabilities, such as gradient boosting decision trees or neural network models, can be used for training using the historical time series dataset prepared in step one to learn the normal causal function relationship under no specific intervention conditions, thereby fitting the functional relationship between each variable and its parent node V_i = f_i(Parents(V_i), U_i), where V_i is variable i, Parents(V_i) is the set of its parent nodes, f_i is the causal mechanism function of variable i, and U_i is the exogenous random disturbance term representing unobserved factors. Next, the system performs counterfactual inference on the constructed SCM for each management intervention event identified in step three. Counterfactual inference is achieved in SCM by executing causal interference operators (i.e., "do" operators).

[0018] Suppose that in the real world, during an audit, management is observed to have implemented an intervention I, with a value of i_actual (e.g., R&D capitalization of 50 million yuan). This results in an observed value of P_actual for a key financial indicator (e.g., net profit). To investigate the true impact of this intervention, the system needs to simulate a counterfactual scenario where management did not implement intervention I. This process is achieved by setting a baseline value i_baseline for the intervention variable I on the SCM (e.g., R&D capitalization of 0), i.e., performing a do(I=i_baseline) operation on the model. After executing the "do" operator, the model cuts off all original causal paths pointing to variable I and forcibly fixes its value to the baseline value i_baseline. Then, based on the causal relationships learned in the SCM, the model recalculates the predicted values ​​of all downstream nodes affected by the intervention, starting from the intervention point and following the transmission path of the financial causal graph, until finally obtaining the predicted value P_counterfactual for the key financial indicator under the counterfactual scenario.

[0019] Step 5: Manipulation Effect Mining. By comparing actual observed values ​​with counterfactual predicted values, the system can accurately identify the manipulation effect value caused by a single intervention. The magnitude of this manipulation effect value is quantified as ΔP = P_actual - P_counterfactual. If the absolute value of ΔP exceeds a pre-set audit materiality level, the system automatically marks the intervention event as a high-risk profit manipulation clue and generates a detailed attribution analysis report. This report may specifically include: a risk level score generated by the system, a detailed description of the intervention behavior, a list of specific accounting items involved, a visualized causal transmission path diagram, precise quantitative impact values ​​on key performance indicators such as net profit and return on assets, and evidence fragments extracted from the original text.

[0020] Furthermore, to enhance the system's robustness and adaptability, in one optional implementation, the machine learning model used to learn the causal functions in the SCM can be upgraded to a probabilistic model capable of uncertainty assessment, such as a Bayesian neural network or a Gaussian process. With such a model, the system can output a confidence interval or uncertainty score in addition to the counterfactual prediction value P_counterfactual. This uncertainty score plays a crucial role: when the system identifies a new or rare management intervention and performs counterfactual analysis, if the uncertainty score of the output result exceeds a preset threshold, it indicates that the current SCM model is insufficient in handling such interventions, and the model may need updating. In this case, the system can trigger an adaptive update process, such as automatically marking the high-uncertainty intervention event and its analysis results, and prompting audit experts for manual review. The expert's review conclusions, such as their professional judgment on the impact of the intervention, can serve as new high-quality labeled data for online learning or incremental fine-tuning of the SCM model, thereby dynamically optimizing and updating the causal relationships in the model.

[0021] Compared with existing technologies, the system and method provided by this invention have the following beneficial effects: This invention constructs a technical process from data acquisition, causal relationship modeling, management intent identification to counterfactual quantitative analysis, to address the limitations of traditional auditing methods that rely on correlation analysis. By constructing a financial causal graph that integrates domain knowledge and data-driven approaches, and utilizing large-scale language models to identify management intervention behaviors, this invention can deepen the perspective of audit analysis from data appearances to the level of business motivations.

[0022] In particular, through counterfactual deduction based on a structured causal model, this invention can isolate and quantify the net impact of a single management decision on financial results, enabling analysis from anomaly detection to causal attribution. This provides auditors with precise and interpretable audit clues, thereby improving the ability to identify complex profit manipulation behaviors hidden beneath compliance appearances, the accuracy of audit judgments, and the overall efficiency of audit work. Furthermore, the technical solution provided by this invention is not only applicable to external audits by certified public accountants, but its application value can also be extended to multiple scenarios such as corporate internal control, investment risk assessment, and financial supervision, providing technical support for different entities to conduct financial condition analysis and business substance judgment. The introduced uncertainty quantification and adaptive update mechanism also enables the system to be dynamically updated to adapt to new business environments and manipulation methods. Attached Figure Description

[0023] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0024] Figure 1This is a schematic diagram of the functional module structure of a financial data causal relationship mining system for intelligent auditing according to the present invention.

[0025] Figure 2 This is a flowchart illustrating a financial data causal relationship mining method for intelligent auditing according to the present invention.

[0026] Figure 3 This is a detailed flowchart illustrating the steps involved in constructing a financial cause-effect graph in an embodiment of the present invention.

[0027] Figure 4 This is a schematic diagram illustrating the principle of manipulation effect mining based on counterfactual deduction in an embodiment of the present invention.

[0028] Figure 5 This is a schematic diagram of an example of a financial causal graph that integrates domain knowledge and data-driven discovery in an embodiment of the present invention.

[0029] Figure 6 This is a schematic diagram illustrating the results of counterfactual reasoning for a specific management intervention event in an embodiment of the present invention.

[0030] Figure 7 This is a schematic diagram of a preferred embodiment of the present invention, which includes the results of uncertainty quantification counterfactual deduction.

[0031] Figure 8 This is a schematic diagram comparing the accuracy of the system of this invention with existing technologies in identifying different types of profit manipulation. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0033] Example 1 Reference Figure 1 This invention provides a financial data causal relationship mining system for intelligent auditing. The system includes at least: a data acquisition and preprocessing module 101, a financial causal graph construction module 102, a management intervention variable identification and quantification module 103, and a counterfactual deduction and manipulation effect mining module 104. Each module can be implemented by software program code in a central processing unit, or by dedicated hardware circuits such as field-programmable gate arrays (FPGAs) or application-specific integrated circuits (ASICs).

[0034] The method provided in this embodiment of the invention can be executed by the above-described system, as described above. Figure 2 The method specifically includes the following steps: Step S201: Perform multi-source heterogeneous financial data acquisition and preprocessing.

[0035] This step establishes a high-quality, standardized data foundation for subsequent causal relationship mining. Specifically, the data acquisition and preprocessing module 101 acquires structured and unstructured data related to the target audit entity from multiple data sources. Structured data may include data from various items in the company's historical annual and quarterly balance sheets, income statements, and cash flow statements, such as operating revenue, operating costs, selling expenses, administrative expenses, accounts receivable, inventory, and fixed assets; it may also include detailed transaction data recorded in the Enterprise Resource Planning (ERP) system, such as sales orders, purchase orders, and inventory changes; simultaneously, to place the company's operations in a broader context, macroeconomic indicators (such as GDP growth rate and consumer price index) and industry average data (such as industry average debt-to-equity ratio and industry average price-to-earnings ratio) may be acquired as external environmental variables. Unstructured data mainly includes management discussion and analysis (MD&A) sections in the company's annual reports, interim announcements regarding major asset restructuring or changes in accounting policies, press releases, industry research reports, and third-party financial news reports—text materials containing a large amount of descriptive language.

[0036] After data acquisition, the data acquisition and preprocessing module 101 performs a series of preprocessing operations. For structured data, data cleaning, imputation, and standardization are performed. For example, for missing quarterly data in financial statements, cubic spline interpolation is preferred for imputation, as this method ensures the smoothness of the time series after imputation, conforming to the gradual change characteristics of financial data. For outliers in the data, unsupervised learning algorithms such as Isolation Forest can be used for identification and correction. This algorithm does not require pre-assumptions about the data distribution and is suitable for complex financial data. Subsequently, all financial item data undergoes caliber alignment and standardization, such as mapping the accounting items of different companies according to unified accounting standards, and Z-score standardization is performed on numerical data to eliminate the influence of units, forming a unified time series dataset.

[0037] For unstructured data, deep natural language processing is performed. This includes sentence segmentation, word segmentation, stop word removal, part-of-speech tagging, and named entity recognition (NER). Specifically, the system employs a financial domain named entity recognition model pre-trained on financial and accounting corpora to accurately identify and extract specific financial items (such as "goodwill" and "deferred tax assets"), accounting policies (such as "revenue recognition timing" and "fixed asset depreciation methods"), key figures (such as "Chief Financial Officer" and "Chairman of the Audit Committee"), and specific economic events (such as "asset restructuring," "equity incentive plans," and "related party transactions") mentioned in the text. The processed text materials form a standardized text corpus. Finally, all pre-processed structured and unstructured data are integrated and stored in a unified data warehouse, providing structured and standardized input for subsequent steps.

[0038] Step S202: Execute the construction of a financial causal graph based on the fusion of domain knowledge and data-driven approaches.

[0039] This step is performed by the financial cause-effect graph construction module 102, which constructs a financial cause-effect graph that describes the causal relationships between various financial indicators, operating activities, and external factors of the enterprise. This financial cause-effect graph is formally a directed acyclic graph (DAG), where nodes represent various variables (such as financial indicators and economic events), and directed edges between nodes represent direct causal relationships between variables. (Refer to...) Figure 3 The specific construction process is as follows: Sub-step S301: Construct an initial causal graph based on domain knowledge. First, based on generally accepted accounting standards, accounting identities, and macroeconomic theories, an initial domain knowledge causal graph is constructed. This initial graph contains generally accepted, deterministic financial causal relationships within the domain. For example, the accounting identity "Assets = Liabilities + Owner's Equity" defines the deterministic relationship between these three macroeconomic variables; business logic determines that "operating revenue" is a direct cause of "operating profit," and "selling expenses" are also a direct cause of "operating profit" (in a negative direction); the principles of cash flow statement preparation determine that "cash paid for the purchase and construction of fixed assets, intangible assets, and other long-term assets" is one of the direct causes of changes in the book value of "fixed assets." This initial graph, constructed based on domain knowledge, provides the initial structure and strong constraints for the entire causal graph construction, effectively avoiding pseudo-causal relationships or directionally incorrect causal relationships that purely data-driven algorithms might learn.

[0040] Sub-step S302 employs a data-driven causal discovery algorithm for learning. Based on the domain knowledge causal graph, a data-driven causal discovery algorithm is used to learn from the structured time series dataset prepared in step S201 to discover hidden, non-obvious potential causal relationships in the data. Preferably, a constraint-based causal discovery algorithm, such as the PC algorithm or FCI algorithm, can be used. These algorithms infer the causal structure between variables by performing a series of conditional independence tests on the data. For example, the algorithm will test whether there is a direct causal relationship between "accounts receivable turnover" and "bad debt provision ratio," or whether their correlation is caused by a common cause (such as "macroeconomic prosperity"). Through analysis of historical data, the algorithm may discover that for a company in a specific industry, changes in "inventory turnover" will significantly affect the ratio of "selling expenses" to revenue in the next accounting period, because a lower turnover may force the company to increase marketing investment to clear inventory. The actual time lag relationship is determined by the model based on data learning.

[0041] Sub-step S303 involves fusing and refining the data to form the final causal graph. Potential causal edges discovered through data-driven discovery undergo statistical significance testing, such as calculating p-values ​​to determine their confidence level. Only causal edges that pass the test are added as candidate edges to the initial knowledge causal graph. During the fusion process, conflict detection is also performed; for example, if a causal relationship discovered by the data-driven algorithm violates a strong constraint in the domain knowledge graph, the data-driven result is discarded. Finally, the system fuses the data-driven, verified causal relationships with the initial knowledge causal graph to form a more complete and accurate financial causal graph, providing a foundation for subsequent causal inference.

[0042] Reference Figure 5 This paper illustrates an example of a financial causal graph constructed according to the method of the present invention. In this example, the system integrates multiple types of causal relationships: firstly, deterministic causal relationships based on knowledge in fields such as accounting standards, such as the causal relationship from "operating revenue" and "operating costs" to "net profit," as shown by the solid lines in the figure; secondly, potential causal relationships learned through data-driven causal discovery algorithms, such as the relationship from "inventory turnover rate" to "selling expenses," as shown by the dashed lines in the figure. In particular, the figure also marks the identified management intervention variable, namely "capitalization of R&D expenditures," with dark gray nodes, and clearly shows the direct impact path of this intervention on "administrative expenses" and "intangible assets" with bold solid lines. This figure provides a structured foundation for subsequent counterfactual reasoning.

[0043] Step S203: Perform management intervention behavior identification and quantification based on a large language model.

[0044] This step is executed by the management intervention variable identification and quantification module 103. Its core function is to automatically identify, from the unstructured text data processed in step S201, accounting decisions or operational behaviors led by management that may have a substantial impact on the company's financial results, and to identify and quantify these behaviors as intervention variables. Preferably, this step employs a large-scale language model that has undergone fine-tuning based on massive amounts of professional financial corpus data, such as a language model based on the Transformer architecture.

[0045] To achieve accurate identification of intervention behaviors, the system pre-constructed a specialized instruction fine-tuning dataset. This dataset contains numerous fragments extracted from listed companies' "management discussions and analyses," announcements, and other texts. Senior audit experts have annotated sentences describing controllable accounting decisions or non-routine business practices. The annotations include the behavior type and key parameters. For example, for the sentence "During the reporting period, the company changed the method for calculating the provision for bad debts on accounts receivable from the aging analysis method to the percentage of balance method. This change in accounting estimate reduced current credit impairment losses by approximately RMB 20 million," the experts would annotate it as the intervention type "Accounting Estimate Change - Bad Debt Provision" and mark the impact parameter "RMB 20 million."

[0046] When performing the recognition task, preprocessed unstructured text is input into the fine-tuned large language model. The model is trained to perform sequence labeling or text classification tasks to identify and extract two types of key information: first, the type of intervention behavior, such as identifying specific behaviors like "accounts receivable factoring," "sale-leaseback transaction," "change in goodwill impairment testing method," or "capitalization of R&D expenditures"; second, the relevant parameters of the intervention behavior, such as quantitative information like the specific amount involved, the time of the transaction, and the counterparty.

[0047] For example, when the model processes the sentence "During this reporting period, the company changed some eligible R&D investments from expense treatment to capitalization treatment, which increased intangible assets by approximately RMB 50 million in the current period," it can automatically identify this as an intervention of type "R&D expenditure capitalization" and extract "RMB 50 million" as the quantitative parameter of this intervention. Finally, the system precisely maps each identified intervention variable to one or more corresponding nodes in the financial causal graph constructed in step S202, serving as intervention points for subsequent counterfactual inference. For example, the "R&D expenditure capitalization" intervention will be mapped to both the "intangible assets" and "administrative expenses" nodes simultaneously.

[0048] Step S204: Perform counterfactual inference based on a structured causal model.

[0049] This step, executed by the counterfactual deduction and manipulation effect mining module 104, is one of the core components of causal attribution. First, the financial causal graph constructed in step S202 is transformed into a structured causal model (SCM). In an SCM, the value of each variable is determined by a function whose inputs are the values ​​of all its direct causes (i.e., parent nodes) in the causal graph, and an exogenous random perturbation term representing unobserved factors. Specifically, the functional relationship between each variable V_i and its parent node set Parents(V_i) can be expressed as V_i = f_i(Parents(V_i), U_i), where f_i is the causal mechanism function of variable i, and U_i is the random perturbation. These causal functions f_i can be learned using machine learning models with strong nonlinear modeling capabilities, such as gradient boosting decision trees or neural network models, and trained using the historical time series dataset prepared in step S201 to learn normal causal relationships under no specific intervention conditions.

[0050] Next, for each management intervention event identified in step S203, the system performs counterfactual reasoning on the constructed SCM. (Refer to...) Figure 4 This process can be understood as comparing two scenarios. In the real world, during the audit, management is observed to have implemented an intervention I, with a value of i_actual (e.g., R&D capitalization of 50 million yuan). This results in an actual observed value of P_actual for key financial indicators (e.g., "net profit"). To investigate the true impact of this intervention, the system needs to simulate a counterfactual scenario—a scenario where management did not implement intervention I. This process is achieved by setting a baseline value i_baseline for the intervention variable I on the SCM (e.g., R&D capitalization of 0).

[0051] Technically, this is accomplished by executing a causal intervention operator, denoted as do(I=i_baseline). After executing the do operator, the model cuts off all original causal paths pointing to variable I and forces its value to be fixed at the baseline value i_baseline. This simulates changing only this one management decision while keeping all other conditions unchanged. Then, based on the causal relationships learned in the SCM, the model recalculates the predicted values ​​of all downstream nodes affected by the intervention, starting from this intervention point and following the transmission path of the financial causal graph, until finally obtaining the predicted value P_counterfactual of the key financial indicator under the counterfactual scenario.

[0052] Step S205: Perform manipulation effect mining.

[0053] This step is executed by the counterfactual deduction and manipulation effect mining module 104. By comparing actual observed values ​​with counterfactual predicted values, the system can accurately mine the manipulation effect value caused by this single intervention behavior. The magnitude of this manipulation effect value is quantified as ΔP = P_actual - P_counterfactual. If the absolute value of ΔP exceeds the pre-set audit materiality level (e.g., 5% of pre-tax profit), the system automatically marks the intervention event as a high-risk profit manipulation clue and generates a detailed attribution analysis report. This report may specifically include: a risk level score generated by the system, a detailed description of the intervention behavior (e.g., "adjusting profits through R&D expenditure capitalization"), a list of specific accounting items involved (e.g., "intangible assets", "administrative expenses", "net profit"), a visualized causal transmission path diagram, and precise quantitative impact values ​​on key performance indicators such as net profit and return on assets (e.g., "this behavior resulted in an increase of RMB 42.5 million in net profit for the current period"), along with evidence fragments extracted from the original text for auditors to directly refer to.

[0054] Reference Figure 6 The study, presented as a time-series line graph, illustrates the counterfactual analysis and manipulation effect analysis of the specific intervention event of "capitalizing 50 million yuan in R&D expenditures." In the graph, after the intervention point in the fourth quarter of 2022 (2022Q4), the curve representing the actual observed net profit (P_actual) and the counterfactual predicted net profit curve (P_counterfactual) representing the net profit without this intervention begin to separate. The shaded area between the two curves visually demonstrates the cumulative manipulation effect of this intervention. By the end of the observation period (2023Q4), this manipulation effect was precisely quantified to 15 million yuan, thus providing auditors with direct, quantitative causal evidence regarding the impact of this management decision on profits.

[0055] In a preferred embodiment, to improve the robustness and adaptability of the system, the machine learning model used to learn the causal function in the SCM can be upgraded to a probabilistic model capable of uncertainty assessment, such as a Bayesian neural network or a Gaussian process. (See also...) Figure 7 This demonstrates the counterfactual inference results, incorporating uncertainty quantification. Using this model, the system outputs a counterfactual prediction value, P_counterfactual, along with an additional confidence interval (shown as the gray area surrounding the P_counterfactual curve in the figure). The width of this confidence interval reflects the model's confidence in the prediction; generally, the further away from the intervention point, the greater the uncertainty, and the wider the interval. When the system identifies a new or rare management intervention and performs a counterfactual inference, if the uncertainty score of the output result exceeds a preset threshold (e.g., the confidence interval is too wide), it indicates that the current SCM model is insufficient to handle such interventions. In this case, the system can trigger an adaptive update process, automatically marking the high-uncertainty intervention event and its analysis results, and prompting audit experts for manual review. The expert's review conclusions can serve as new high-quality labeled data for online learning or incremental fine-tuning of the SCM model, thereby dynamically optimizing and updating the causal relationships in the model, enabling the system to adapt to new business environments and manipulation methods.

[0056] To verify the beneficial effects of the system and method of the present invention, refer to Figure 8 This paper compares the accuracy of the system of this invention with traditional correlation analysis-based auditing methods in identifying four typical profit manipulation behaviors. These four behaviors, arranged in ascending order of difficulty, are: inflated revenue, related-party transactions, capitalization of R&D expenditures, and changes in accounting estimates. As shown in the figure, both methods achieve high accuracy for direct financial fraud such as inflated revenue. However, for deeper, compliant profit manipulation behaviors such as capitalization of R&D expenditures and changes in accounting estimates, which are hidden beneath compliant accounting policy choices, the accuracy of traditional correlation analysis methods drops significantly to 40% and 30%, respectively. In contrast, the system of this invention, employing a causal inference analysis path, can penetrate the surface of compliance and maintains high accuracy rates of 88% and 85%, respectively. This comparison demonstrates the technical effectiveness of this invention in improving the identification of complex and concealed profit manipulation behaviors.

[0057] Example 2 This embodiment provides a financial data causal relationship mining system for intelligent auditing, the structure of which is as follows: Figure 1 As shown. The system includes at least: The data acquisition and preprocessing module 101 is used to perform step S201 as described in Embodiment 1, to acquire structured and unstructured data from multiple data sources, and to perform data cleaning, standardization and natural language processing operations, so as to establish a high-quality and standardized data foundation for subsequent modules.

[0058] The financial causal graph construction module 102 is used to execute step S202 as described in Embodiment 1. By integrating knowledge from fields such as accounting standards with data-driven causal discovery algorithms, it constructs a financial causal graph that describes the causal relationships between various financial indicators, operating activities and external factors of an enterprise.

[0059] The management intervention variable identification and quantification module 103 is used to perform step S203 as described in Embodiment 1. It uses a large language model fine-tuned in the financial field to automatically identify and quantify management intervention variables and their intervention parameters that are led by management and may affect financial results from unstructured text data.

[0060] The counterfactual deduction and manipulation effect mining module 104 is used to perform step S204 as described in Embodiment 1. First, the financial causal diagram is transformed into a structured causal model. Then, for the identified management intervention events, counterfactual deduction is performed to quantify the net impact of the intervention on key financial indicators, i.e., the manipulation effect, and an audit risk clue is generated based on its magnitude.

[0061] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for mining causal relationships in financial data for intelligent auditing, characterized in that, Includes the following steps: Acquire multi-source heterogeneous data of the target audit entity, including structured financial data and unstructured text data, and perform preprocessing on the structured financial data and the unstructured text data to form a unified time series dataset and standardized text corpus; By integrating domain knowledge with data-driven causal discovery algorithms, the final financial causal graph is constructed. Based on the unified time series dataset, a machine learning model is used to fit a causal function with its parent node as input to each variable in the final financial causal graph, thereby instantiating the final financial causal graph into a structured causal model. The standardized text corpus is processed using a large language model fine-tuned with instructions from the financial domain corpus. The management intervention behavior is identified through sequence labeling and text classification. The corresponding intervention type and intervention parameters are extracted, and the intervention variables corresponding to the management intervention behavior in the structured causal model are determined. For each identified management intervention behavior, counterfactual inference is performed in the structured causal model. Specifically, the causal path pointing to the corresponding intervention variable is cut off, its value is fixed to a preset benchmark value, and the predicted value of the downstream financial indicator is recalculated according to the causal function to obtain the counterfactual predicted value. Calculate the difference between the actual observed value and the counterfactual predicted value of at least one financial indicator, quantify the difference as a manipulation effect value, and generate audit risk clues based on the manipulation effect value.

2. The method according to claim 1, characterized in that, The construction of the final financial causal graph includes: constructing an initial domain knowledge causal graph based on generally accepted accounting principles and accounting identities; using a data-driven causal discovery algorithm to learn the unified time series dataset to discover potential causal relationships hidden in the data; determining whether the potential causal relationships are logically compatible with the initial domain knowledge causal graph, and adding logically compatible potential causal relationships that pass the statistical significance test as new edges to the initial domain knowledge causal graph to form the final financial causal graph.

3. The method according to claim 1, characterized in that, The preprocessing of the structured financial data includes: using cubic spline interpolation to handle missing values, and using the isolated forest algorithm to identify and correct outliers; the preprocessing of the unstructured text data includes: using a pre-trained financial domain named entity recognition model to identify and extract specific financial items, accounting policies, and economic events mentioned in the text.

4. The method according to claim 1, characterized in that, The data-driven causal discovery algorithm is a constraint-based causal discovery algorithm, including the Peter-Clark algorithm or the Fast Causal Inference (FCI) algorithm.

5. The method according to claim 1, characterized in that, The large-scale language model is a Transformer-based language model that has been fine-tuned based on instructions on a massive amount of professional financial corpus. The process of identifying management intervention behaviors and extracting their corresponding intervention types and parameters is as follows: the standardized text corpus is input into the large-scale language model, and the model performs sequence labeling or text classification tasks to identify and extract the type of intervention behavior and related parameters of the intervention behavior. The related parameters of the intervention behavior include the amount involved in the intervention behavior, the time of occurrence, or the counterparty of the transaction.

6. The method according to claim 5, characterized in that, The instruction fine-tuning was accomplished based on a pre-built instruction fine-tuning dataset, which includes multiple samples selected from the "Management Discussion and Analysis" text and contains annotations by audit experts on sentences describing controllable accounting decisions or non-routine business practices.

7. The method according to claim 1, characterized in that, The machine learning model is a gradient boosting decision tree model or a neural network model.

8. The method according to claim 7, characterized in that, The machine learning model is a probabilistic model capable of uncertainty assessment, including a Bayesian neural network or a Gaussian process model; the counterfactual inference outputs a confidence interval or uncertainty score along with the counterfactual prediction value.

9. The method according to claim 8, characterized in that, The method also includes an adaptive update process. When the uncertainty score is higher than a preset threshold, the adaptive update process is triggered to mark the management intervention behavior and prompt the audit expert to conduct a manual review. The audit expert's review conclusion is used as new labeled data for online learning or incremental fine-tuning of the structured causal model.

10. A financial data causal relationship mining system for intelligent auditing, characterized in that, include: The data acquisition and preprocessing module is used to acquire structured financial data and unstructured text data of the target audit entity, and to preprocess them to form a unified time series dataset and standardized text corpus. The financial causal graph construction module is used to construct an initial domain knowledge causal graph based on generally accepted accounting principles and accounting identities, and integrate the potential causal relationships learned through data-driven causal discovery algorithms and tested for logical compatibility and statistical significance to form the final financial causal graph. The module for identifying and quantifying management intervention variables is used to identify management intervention behaviors and their parameters from the standardized text corpus using a large language model. The counterfactual deduction and manipulation effect mining module is used to instantiate the final financial causal graph into a structured causal model based on the unified time series dataset, and to perform counterfactual deduction on the management intervention behavior, calculate the manipulation effect value, and generate audit risk clues.

Images