Enterprise financial management risk prediction method based on big data analysis

Abstract

The application discloses an enterprise financial management risk prediction method based on big data analysis, comprising the following steps: collecting and processing financial basic data to generate standard financial time series data; constructing a financial PatchTST model; identifying a financial event trigger point based on the standard financial time series data to generate a financial event trigger sequence; generating a dynamic financial Patch sequence based on the financial event trigger sequence; organizing subject group Patch embedding representation based on the dynamic financial Patch sequence to generate a subject group financial embedding sequence; generating a financial risk conduction coding representation based on the subject group financial embedding sequence; generating an enterprise financial management risk prediction result based on the financial risk conduction coding representation; and generating financial risk warning information based on the enterprise financial management risk prediction result, the financial event trigger sequence and the financial risk conduction coding representation. The application improves prediction accuracy, interpretability and timeliness.

Description

Technical Field

[0001] This invention relates to the field of corporate financial risk management, and in particular to a method for predicting corporate financial management risks based on big data analysis. Background Technology

[0002] As businesses expand and their systems become more digitalized, the amount of data generated in corporate financial management—including accounting balances, transaction data, accounts receivable aging data, accounts payable due data, cash flow data, revenue data, expense data, and liability data—continues to grow. This data is characterized by multiple sources, inconsistent periods, significant differences in account definitions, and long cash flow chains. Traditional methods of financial risk identification typically rely on manual rules, fixed threshold indicators, and post-event report analysis, making it difficult to reflect the continuous impact of accounts receivable ties, cash flow pressure, concentrated payments, and profit erosion in a timely manner.

[0003] Existing methods for predicting corporate financial risk often employ static financial indicator calculations, single time series models, and general deep learning models for risk identification. These methods typically segment financial data into fixed time windows, failing to adjust the analysis boundaries based on the location of financial events. This can weaken the impact of key events such as abnormal aging migration, concentrated transactions, and widening cash flow gaps during the prediction process. Furthermore, existing methods underutilize the relationships between accounting items, debit / credit directions, cash flow patterns, and accounts receivable maturity dates, making it difficult for the models to express the transmission process of risk between different account groups.

[0004] In addition, existing forecasts typically only output risk scores or risk levels, lacking a correlation with risk source data, risk trigger locations, risk transmission paths, and the time frame of risk occurrence. The interpretability of the early warning information is insufficient, making it difficult to support corporate financial managers in tracing and handling risks.

[0005] Therefore, how to provide a method for predicting corporate financial management risks based on big data analysis is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a method for predicting corporate financial management risks based on big data analysis. This invention constructs an excitation-based financial PatchTST model based on standard financial time-series data. It generates financial event excitation sequences through a financial event excitation unit, dynamic financial patch sequences through a dynamic financial patch construction unit, and financial embedding sequences for account groups through an account group patch embedding unit. Furthermore, it generates a financial risk transmission encoding representation through an excitation-based attention encoding unit, and then generates corporate financial management risk prediction results through a phased progressive prediction unit. This results in financial risk early warning information that includes risk level, risk source data, risk trigger location, risk transmission path, and risk occurrence time range. This method possesses the advantages of accurate risk identification, clear risk transmission, and complete early warning information.

[0007] A method for predicting enterprise financial management risks based on big data analysis according to an embodiment of the present invention includes the following steps: Collect basic financial data, perform field cleaning, time alignment, account mapping and missing value filling to generate standard financial time series data; Construct a financial PatchTST model, including a financial event triggering unit, a dynamic financial Patch construction unit, a subject group Patch embedding unit, a triggering and transmission attention encoding unit, and a stage-progressive prediction unit; Based on standard financial time-series data, financial event triggering points are identified in the financial event triggering unit, and a financial event triggering sequence is generated. Based on the financial event triggering sequence, standard financial time series data is converted into financial event patches around the financial event triggering point through dynamic financial patch construction units, generating a dynamic financial patch sequence. Based on the dynamic financial patch sequence, the relationship between account groups is constructed through the account group patch embedding unit, and the dynamic financial patch sequence is organized into an account group patch embedding representation to generate the account group financial embedding sequence. Based on the financial embedding sequence of subject groups, the transmission relationship of financial event triggering points in the subject group relationship is encoded by stimulating the transmission attention encoding unit, thereby generating a financial risk transmission encoding representation; Based on the financial risk transmission coding representation, risk prediction is performed through a phased progressive prediction unit to generate enterprise financial management risk prediction results; Based on the enterprise financial management risk prediction results, financial event triggering sequence and financial risk transmission code representation, financial risk early warning information is generated.

[0008] Optionally, the generation of standard financial time-series data includes: Collect basic financial data, including accounting balance data, transaction flow data, accounts receivable aging data, accounts payable due data, cash flow data, revenue data, expense data, and liability data, and write them into the same data processing space according to the data source identifier to form the basic financial data to be cleaned; Based on the financial basic data to be cleaned, perform field cleaning to generate cleaned financial basic data; Based on the cleaned financial data, time alignment is performed, and data is written to the same time position to generate time-aligned financial data. Based on time-aligned financial basic data, perform account mapping, write it under the corresponding unified account code, and generate account mapping financial basic data; Based on the financial basic data of account mapping, identify the location of missing data, perform missing value filling, and generate the filled financial basic data; Based on the supplemented financial data, fields are combined to generate standard financial time-series data.

[0009] Optionally, the construction of the financial PatchTST model includes: The basic model structure is built based on the PatchTST model, including the Patch partitioning structure, Patch embedding structure, positional encoding structure, Transformer encoding structure, and prediction head structure. Based on the basic model structure, a financial event triggering unit is added; Based on the Patch partitioning structure, the partitioning method of forming Patch input markers according to a fixed time length is replaced by a dynamic financial Patch construction method that adjusts the boundaries of continuous time segments according to the financial event triggering sequence, and a dynamic financial Patch construction unit is constructed. Based on the Patch embedding structure, the processing method of Patch embedding on a single data channel is replaced by the processing method of Patch embedding on dynamic financial Patch sequences based on the relationship of account groups, and account group Patch embedding units are constructed. Based on the Transformer encoding structure, the self-attention connection structure is replaced with an attention connection structure that introduces the relationship between the financial event triggering sequence and the subject group, and a triggering and conduction attention encoding unit is constructed. Based on the forecast head structure, the single output structure that directly outputs the time series forecast results is replaced with a forecast structure that outputs progressively according to the status of accounts receivable occupancy risk, cash flow pressure risk, repayment concentration risk, profit erosion risk and comprehensive financial management risk, thus constructing a stage-progressive forecast unit. Based on the connection relationship between the financial event triggering unit, the dynamic financial patch construction unit, the subject group patch embedding unit, the triggering transmission attention encoding unit, and the stage-progressive prediction unit, the triggering financial patchTST model is generated.

[0010] Optionally, the generation of the financial event triggering sequence includes: Based on standard financial time-series data, in the financial event triggering unit, based on the accounting account balance data, adjacent data values ​​are compared according to the time index to determine the position of the accounting account balance change. Based on transaction flow data, the number of transactions, total transaction amount, and concentration of transaction objects are statistically analyzed to determine the location of concentrated transaction flow. Based on accounts receivable aging data, the aging levels are compared according to the time index to determine the aging migration position; Based on accounts payable due date data, the due date is matched to determine the due date position of accounts payable; Based on cash flow data, extract the values ​​of operating cash inflows and operating cash outflows to determine the cash flow gap and identify the location where the cash flow gap is widening. Based on the location of changes in accounting account balances, the location of concentrated transaction flows, the location of aging migration, the location of accounts payable due, and the location of widening cash flow gaps, the location is merged in the financial event triggering unit. The time index after location merging is determined as the financial event triggering point, and the corresponding financial event triggering intensity is generated. Based on the trigger points and intensity of financial events, financial event trigger sequences are generated by arranging the trigger points in the time order of standard financial time series data.

[0011] Optionally, generating the dynamic financial patch sequence includes: Based on standard financial time-series data, in the dynamic financial patch construction unit, continuous time segments are divided according to the time index sequence and the basic segmentation length, and the merging interval threshold and the number of overlapping time indices are determined. Based on the financial event trigger sequence, the time index corresponding to the financial event trigger point is matched with the time index range corresponding to the continuous time segment to generate the continuous time segment matching result. Based on the continuous time segment matching results, select the continuous time segment containing the financial event trigger point, and determine the time index corresponding to the financial event trigger point as the financial event patch center; Based on the financial event patch center, the boundaries of continuous time segments are adjusted and merged into financial event patches; Based on the continuous time segment matching results, continuous time segments that do not contain financial event trigger points are selected and converted into basic financial patches. A dynamic financial patch sequence is generated based on financial event patches and basic financial patches.

[0012] Optionally, the generated account group financial embedding sequence includes: Based on the dynamic financial patch sequence, in the subject group patch embedding unit, the accounting subject balance data, transaction flow data, accounts receivable aging data, accounts payable due data, cash flow data, revenue data, expense data and liability data of each financial event patch and the basic financial patch are extracted. Based on the unified account code and the superior account code in the accounting account balance data, the account affiliation relationship is determined, account group numbers are assigned, and the debit and credit direction relationship is determined based on the debit and credit amounts. Based on the payer, payee, transaction amount and transaction time in the transaction log data, determine the relationship of fund flow; Based on the time sequence of accounts receivable aging data, accounts payable due data, and standard financial time series data, the due date relationships are determined. Based on the relationships of account affiliation, debit and credit direction, fund flow, and accounts receivable maturity, the data in the dynamic financial patch sequence is divided into embedded channels according to the account group number, and a relationship identifier vector is generated. Based on the embedded channels and relational identifier vectors, accounting account balance data, transaction flow data, accounts receivable aging data, accounts payable due data, cash flow data, revenue data, expense data, and liability data are converted into financial data value vectors, and then combined with the corresponding relational identifier vectors to generate the account group Patch embedded representation, which is then arranged to generate the account group financial embedded sequence.

[0013] Optionally, the generation of the financial risk transmission code representation includes: Based on the financial embedding sequence of subject group, the subject group Patch embedding representation under the same time index is used as the attention encoding object to form the embedding sequence to be encoded. Based on the subject group association relationship, the subject group Patch embedding representation in the embedding sequence to be encoded is connected and constrained, and connectable and non-connectable objects are marked to generate an attention connection mask. Based on the borrowing and lending direction relationship and fund flow relationship in the subject group association relationship, directional constraints are set on connectable objects to generate connectable objects with directional constraints; Based on the financial event triggering sequence, the financial event triggering intensity is written into a connectable object with directional constraints, the encoding weights are adjusted, and a triggering weighted connectable object is generated. Based on the excitation weighted connection object, attention encoding is performed on the embedded sequence to be encoded in the excitation transmission attention encoding unit to generate the transmission relationship of financial event excitation points in the account group relationship; Based on the transmission relationship, generate risk transmission paths; Based on the transmission relationship and risk transmission path, the financial event trigger point, coding weight, directional constraint, trigger weighted connection object and time sequence in the financial event trigger sequence are combined to generate a financial risk transmission coding representation.

[0014] Optionally, the generation of enterprise financial management risk prediction results includes: Based on the financial risk transmission coding representation, the phased progressive forecasting unit is set up with the following stages: accounts receivable occupancy forecasting stage, cash flow pressure forecasting stage, repayment concentration forecasting stage, profit erosion forecasting stage, and comprehensive risk forecasting stage. Based on the financial risk transmission coding representation, the coding content corresponding to the accounts receivable aging data and revenue data is extracted in the accounts receivable occupancy prediction stage to generate the accounts receivable occupancy risk status. Based on the financial risk transmission coding representation and the accounts receivable occupancy risk status, the coding content corresponding to cash flow data and transaction flow data is extracted in the cash flow pressure forecasting stage to generate the cash flow pressure risk status. Based on the financial risk transmission coding representation and the cash flow pressure risk status, the coding content corresponding to the accounts payable maturity data and liability data is extracted in the solvency concentration forecasting stage to generate the solvency concentration risk status. Based on the financial risk transmission coding representation and the solvency concentration risk status, the coding content corresponding to the revenue and expense data is extracted in the profit erosion prediction stage to generate the profit erosion risk status. Based on the financial risk transmission code representation, accounts receivable occupancy risk status, cash flow pressure risk status, solvency concentration risk status, and profit erosion risk status, a progressive summary is performed in the comprehensive risk prediction stage to generate a comprehensive financial management risk status. Based on the risk status of accounts receivable ties, cash flow pressure, payment concentration, profit erosion, and comprehensive financial management risk, a combination of risk status is used to generate a prediction result of corporate financial management risk.

[0015] Optionally, the generation of financial risk warning information includes: Based on the enterprise financial management risk prediction results, the predicted values ​​corresponding to the status are extracted and divided into low risk level, medium risk level, high risk level and severe risk level to generate risk level; Based on the risk status with the highest predicted value in the enterprise financial management risk prediction results, determine the financial basic data corresponding to the risk status and generate risk source data; Based on the financial event trigger sequence, extract the financial event trigger points corresponding to the risk source data and determine them as risk trigger locations; Based on the financial risk transmission code representation, the risk transmission path corresponding to the risk triggering location is extracted and used as the risk transmission path field in the financial risk early warning information; Based on the risk trigger location and risk transmission path fields, extract the risk occurrence start time index and risk occurrence end time index to determine the risk occurrence time range; Based on risk level, risk source data, risk trigger location, risk transmission path fields, and risk occurrence time range, financial risk warning information is generated by combining fields according to the same risk status.

[0016] The beneficial effects of this invention are: First, this invention constructs a financial PatchTST model based on standard financial time-series data. By using financial event triggering units, it identifies financial event triggering points according to the location of changes in accounting item balances, the location of concentrated transaction flows, the location of aging migration, the location of accounts payable due, and the location of widening cash flow gaps. It then generates a financial event triggering sequence. Compared with the method of risk judgment based solely on a single financial indicator threshold, this invention can extract the key abnormal change locations in the process of corporate financial management risk formation in advance, improve the pertinence and timeliness of risk identification, and reduce the omissions caused by the dispersion of financial data fluctuations and the lag of risk characteristics.

[0017] Secondly, this invention transforms standard financial time-series data into financial event patches centered around the triggering point of a financial event through a dynamic financial patch construction unit, and generates a dynamic financial patch sequence. This overcomes the shortcomings of traditional fixed time segment division methods, which are difficult to adapt to sudden changes in financial risks. It enables the model to extract risk change characteristics in consecutive time segments before and after the triggering point of a financial event, enhances the expressive ability of accounts receivable migration, cash flow gap expansion, and the formation of concentrated repayment pressure, and improves the accuracy and stability of corporate financial management risk prediction results.

[0018] Furthermore, this invention generates a financial embedding sequence of account groups by using an account group patch embedding unit to generate account group financial embedding sequences according to account affiliation, debit / credit direction, cash flow, and accounts receivable maturity. It also generates a financial risk transmission code representation by stimulating the transmission attention encoding unit. Then, the stage-progressive prediction unit performs risk prediction in a progressive order of accounts receivable occupancy risk status, cash flow pressure risk status, repayment concentration risk status, profit erosion risk status, and comprehensive financial management risk status. This enables financial risk warning information to simultaneously include risk level, risk source data, risk trigger location, risk transmission path, and risk occurrence time range, improving the interpretability of warning results and business positioning efficiency. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of a corporate financial management risk prediction method based on big data analysis proposed in this invention; Figure 2 This is a schematic diagram of the financial event triggering sequence and dynamic financial patch sequence generation process proposed in this invention; Figure 3 This is a schematic diagram of the process for generating enterprise financial management risk prediction results, based on the subject group financial embedding sequence, financial risk transmission coding representation, and proposed in this invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0021] refer to Figures 1-3 A method for predicting corporate financial management risks based on big data analysis includes the following steps: Collect basic financial data, perform field cleaning, time alignment, account mapping and missing value filling on the basic financial data to generate standard financial time series data. The basic financial data includes accounting account balance data, transaction flow data, accounts receivable aging data, accounts payable due data, cash flow data, revenue data, expense data and liability data. A financial PatchTST model is constructed, which includes a financial event triggering unit, a dynamic financial Patch construction unit, a subject group Patch embedding unit, a triggering and attention encoding unit, and a stage-progressive prediction unit. Based on standard financial time series data, in the financial event triggering unit, financial event triggering points are identified according to the position of changes in accounting account balances, the position of concentrated transaction flow, the position of aging migration, the position of accounts payable due, and the position of widening cash flow gap. Financial event triggering sequence is then generated according to the time order of the financial event triggering points in the standard financial time series data. Based on the financial event triggering sequence, the standard financial time series data is divided into continuous time segments by constructing a dynamic financial patch. The boundaries of the continuous time segments are adjusted, and the continuous time segments are converted into financial event patches formed around the financial event triggering point. The dynamic financial patch sequence is generated according to the time order of the financial event patches in the standard financial time series data. Based on the dynamic financial patch sequence, the account group relationship is constructed by using the account group patch embedding unit according to the account affiliation, debit / credit direction, cash flow and accounts due date. Based on the account group relationship, the accounting account balance data, transaction flow data, accounts receivable aging data, accounts payable due date data, cash flow data, revenue data, expense data and liability data in the dynamic financial patch sequence are organized into account group patch embedding representations. The account group financial embedding sequence is generated according to the time sequence corresponding to the account group patch embedding representations. Based on the financial embedding sequence of account groups, the financial event triggering sequence and the relationship between account groups, the transmission relationship of financial event triggering points in the relationship between account groups is encoded by the triggering transmission attention encoding unit, and a financial risk transmission encoding representation is generated. Based on the financial risk transmission coding representation, risk prediction is performed in a progressive order of accounts receivable pressure risk, cash flow pressure risk, solvency concentration risk, profit erosion risk and comprehensive financial management risk through a phased progressive prediction unit, generating enterprise financial management risk prediction results. Based on the enterprise financial management risk prediction results, financial event triggering sequence and financial risk transmission code representation, financial risk early warning information is generated. The financial risk early warning information includes risk level, risk source data, risk trigger location, risk transmission path and risk occurrence time range.

[0022] In this embodiment, generating standard financial time-series data includes: Financial basic data is collected based on the enterprise number, including accounting subject balance data, transaction flow data, accounts receivable aging data, accounts payable due data, cash flow data, revenue data, expense data and liability data, and written into the same data processing space according to the data source identifier to form financial basic data to be cleaned; Based on the financial basic data to be cleaned, field cleaning is performed to delete duplicate records with the same enterprise number, time index, account name, and data value. The unit of amount in the accounting account balance data, transaction flow data, accounts receivable aging data, accounts payable due data, cash flow data, revenue data, expense data, and liability data is unified to the same unit of amount. Debit entries are recorded as positive values ​​and credit entries are recorded as negative values, generating cleaned financial basic data. Based on the cleaned financial basic data, time alignment is performed. Accounting balance data, transaction flow data, accounts receivable aging data, accounts payable due data, cash flow data, revenue data, expense data and liability data are merged according to daily time index. Data values ​​corresponding to the same enterprise number, the same daily time index, the same account code and the same financial data type are written to the same time position to generate time-aligned financial basic data. Based on time-aligned financial basic data, an account mapping is performed, matching the account names in different sets of accounts according to a unified account code. The accounting account balance data, transaction flow data, accounts receivable aging data, accounts payable due data, cash flow data, revenue data, expense data, and liability data are written into the corresponding unified account code to generate account mapping financial basic data. Based on the financial basic data of account mapping, the location of missing data is identified. The location of missing data is determined by the enterprise number, daily time index, unified account code and financial data type, and a missing data location identifier is generated. Missing value imputation is performed based on the financial basic data of the account mapping and the missing data location identifier. For missing data locations with adjacent daily time index data values ​​for the same enterprise number, the same unified account code, and the same financial data type, the adjacent daily time index data values ​​are used for imputation. For missing data locations without adjacent daily time index data values ​​but with historical data values ​​of the same financial data type, the average result of historical data values ​​of the same financial data type is used for imputation. For missing data locations with business occurrence records in the transaction flow data, the transaction amount in the business occurrence record is used for imputation, generating the imputed financial basic data. Based on the supplemented financial basic data, the enterprise number, daily time index, unified account code, financial data type and data value are combined into fields to generate standard financial time series data according to the order of the daily time index.

[0023] In this embodiment, constructing the financial PatchTST model includes: The basic model structure is built based on the PatchTST model. The basic model structure includes a Patch partitioning structure, a Patch embedding structure, a positional encoding structure, a Transformer encoding structure, and a prediction head structure. The Patch partitioning structure is used to divide the time series into Patch input labels. The Patch embedding structure is used to convert the Patch input labels into embedded representations. The positional encoding structure is used to write the positional order of the Patch input labels in the time series. The Transformer encoding structure is used to perform temporal encoding on the embedded representations. The prediction head structure is used to output the time series prediction results. Based on the basic model structure, a financial event triggering unit is added. Standard financial time series data is used as the input of the financial event triggering unit, and the output of the financial event triggering unit is connected to the dynamic financial patch construction unit to form an event triggering processing structure that is located before the dynamic financial patch construction unit. Based on the Patch partitioning structure, the partitioning method of forming Patch input markers according to a fixed time length in the Patch partitioning structure is replaced with a dynamic financial Patch construction method that adjusts the boundaries of continuous time segments according to the financial event triggering sequence. A dynamic financial Patch construction unit is constructed. The input end of the dynamic financial Patch construction unit is connected to the standard financial time series data and the financial event triggering unit, respectively. The output end of the dynamic financial Patch construction unit is connected to the subject group Patch embedding unit. Based on the Patch embedding structure, the processing method of Patch embedding a single data channel in the Patch embedding structure is replaced by the processing method of Patch embedding a dynamic financial Patch sequence based on the relationship of subject group. Subject group Patch embedding unit is constructed. The input of subject group Patch embedding unit is connected to dynamic financial Patch construction unit, and the output of subject group Patch embedding unit is connected to the stimulation conduction attention encoding unit. Based on the Transformer coding structure, the self-attention connection structure in the Transformer coding structure is replaced with an attention connection structure that introduces the relationship between financial event triggering sequence and subject group, and a triggering and transmission attention coding unit is constructed. The input of the triggering and transmission attention coding unit is connected to the subject group patch embedding unit, the financial event triggering unit and the processing position in the subject group patch embedding unit that forms the relationship between subject groups, respectively. The output of the triggering and transmission attention coding unit is connected to the stage progressive prediction unit. Based on the forecast head structure, the single output structure that directly outputs the time series forecast results is replaced with a forecast structure that outputs progressively according to the status of accounts receivable occupancy risk, cash flow pressure risk, solvency concentration risk, profit erosion risk, and comprehensive financial management risk. A stage-progressive forecast unit is constructed. The input end of the stage-progressive forecast unit is connected to the stimulation and transmission attention encoding unit, and the output end of the stage-progressive forecast unit is used to output the enterprise financial management risk forecast results. Based on the connection relationship between the financial event triggering unit, the dynamic financial patch construction unit, the subject group patch embedding unit, the triggering transmission attention encoding unit and the stage-progressive prediction unit, the triggering financial patchTST model is generated. When training the financial PatchTST model, historical financial basic data, historical accounts receivable overdue records, historical cash flow gap records, historical payment delay records, historical profit anomaly records, historical comprehensive risk rating records, and historical risk disposal records are collected and merged according to the same enterprise number, the same time index, and the same unified account code to form a historical training sample set. The historical training sample set is then subjected to the same field cleaning, time alignment, account mapping, and missing value imputation processing as the financial basic data to generate historical standard financial time series data. Based on historical standard financial time-series data, historical overdue accounts records are marked as accounts receivable entrapment risk status labels, historical cash flow gap records are marked as cash flow pressure risk status labels, historical payment delay records are marked as payment concentration risk status labels, historical profit abnormality records are marked as profit erosion risk status labels, and historical comprehensive risk rating records are marked as comprehensive financial management risk status labels; the trigger time index in historical risk disposal records is marked as risk trigger location label, the risk transmission record in historical risk disposal records is marked as risk transmission path label, and the start and end time of risk occurrence in historical risk disposal records is marked as risk occurrence time range label; Historical standard financial time-series data is input into the PatchTST financial model, which sequentially passes through the financial event triggering unit, dynamic financial patch construction unit, account group patch embedding unit, triggering transmission attention encoding unit, and stage-progressive prediction unit. The model outputs prediction results for the risk status of accounts receivable occupancy, cash flow pressure, solvency concentration, profit erosion, comprehensive financial management, risk trigger location, risk transmission path, and risk occurrence time range. The training loss consists of risk state prediction loss, risk trigger location loss, risk propagation path loss, and risk occurrence time range loss. Risk state prediction loss is calculated from the cross-entropy between the five risk state prediction results and their corresponding risk state labels. Risk trigger location loss is calculated from the cross-entropy between the risk trigger location prediction results and their risk trigger location labels. Risk propagation path loss is calculated from the cross-entropy between the risk propagation path prediction results and their risk propagation path labels. Risk occurrence time range loss is calculated from the mean absolute error between the risk occurrence time range prediction results and their risk occurrence time range labels. The total training loss is obtained by weighting and summing these losses according to preset weights. The Adam optimizer is used to iteratively train the stimulated financial PatchTST model. The training batch size, learning rate, and training epochs are determined based on the data size of the historical training sample set, the time index length of the standard financial time series data, and the average number of dynamic financial patch sequences. After each training epoch, the validation loss is calculated on the validation sample set. When the decrease in validation loss over multiple consecutive training epochs is less than the preset convergence range, or when the training epochs reach the upper limit, training is stopped and the model parameters corresponding to the minimum validation loss are saved, resulting in the trained stimulated financial PatchTST model.

[0024] In this embodiment, generating the financial event trigger sequence includes: Based on standard financial time series data, in the financial event triggering unit, based on the accounting account balance data corresponding to the same enterprise number and the same unified account code, adjacent data values ​​are compared according to the time index, and the position of accounting account balance change is determined according to the change range, change direction and number of continuous changes between adjacent data values. Based on standard financial time series data, in the financial event triggering unit, based on the transaction flow data corresponding to the same enterprise number and the same time index, the number of transactions, the total transaction amount, and the concentration of transaction objects are counted, and the concentrated position of transaction flow is determined based on the number of transactions, the total transaction amount, and the concentration of transaction objects. Based on standard financial time series data, in the financial event triggering unit, based on the accounts receivable aging data corresponding to the same enterprise number and the same unified account code, the aging level is compared according to the time index, and the time index of the accounts receivable aging data from the short aging level to the long aging level is determined as the aging migration position. Based on standard financial time series data, in the financial event triggering unit, based on the accounts payable due date data corresponding to the same enterprise number and the same unified account code, the due date in the accounts payable due date data is matched, the due date in the accounts payable due date data is mapped to the time index in the standard financial time series data, and the mapped time index is determined as the accounts payable due date position. Based on standard financial time series data, in the financial event triggering unit, based on the cash flow data corresponding to the same enterprise number and the same time index, the values ​​of operating cash inflow data and operating cash outflow data are extracted. The difference between the value of operating cash outflow data and the value of operating cash inflow data is determined as the cash flow gap value. The location of the cash flow gap expansion is determined according to the direction of change and the number of consecutive expansions of the cash flow gap value. Based on the location of changes in accounting account balances, the location of concentrated transaction flows, the location of aging migration, the location of accounts payable due, and the location of widening cash flow gaps, the financial event triggering unit is merged according to the enterprise number and time index. The time index after location merging is determined as the financial event triggering point. Based on the magnitude of change, the number of continuous changes, the number of location merging, and the number of historical risk correspondences recorded in the trained financial event triggering unit, the financial event triggering intensity corresponding to the financial event triggering point is generated. Based on the trigger points and intensity of financial events, the trigger points are arranged in the financial event triggering unit according to their time order in the standard financial time series data. The enterprise number, time index, unified account code, financial data type and intensity of each financial event trigger point are written into the same sequence position to generate a financial event triggering sequence.

[0025] In this embodiment, generating a dynamic financial patch sequence includes: Based on standard financial time series data, in the dynamic financial patch construction unit, continuous time segments are divided according to the order of time indexes and the basic segment length. Half of the basic segment length is determined as the merging interval threshold, and one-quarter of the basic segment length is determined as the number of overlapping time indices. Based on the financial event triggering sequence, the time index corresponding to the financial event triggering point is matched with the time index range corresponding to the continuous time segment in the dynamic financial patch construction unit to generate the continuous time segment matching result. Based on the continuous time segment matching results, select the continuous time segment containing the financial event trigger point, and determine the time index corresponding to the financial event trigger point as the financial event patch center. When the same continuous time segment contains multiple financial event trigger points, determine the time index corresponding to the financial event trigger point with the strongest financial event trigger intensity as the financial event patch center. Based on the financial event patch center, the boundaries of continuous time segments are adjusted. The continuous time segments adjacent to the financial event patch center, the continuous time segment where the financial event patch center is located, and the continuous time segments adjacent to the financial event patch center are merged into a financial event patch. The adjacent time index between adjacent financial event patches is retained according to the number of overlapping time indexes. Based on the merging interval threshold, the time index interval between adjacent financial event trigger points is judged. When the time index interval between adjacent financial event trigger points is less than the merging interval threshold, the financial event patches corresponding to the adjacent financial event trigger points are merged into the same financial event patch. When the time index interval between adjacent financial event trigger points reaches the merging interval threshold, the financial event patches corresponding to the adjacent financial event trigger points are retained. Based on the continuous time segment matching results, select continuous time segments that do not contain financial event trigger points, and convert the continuous time segments that do not contain financial event trigger points into basic financial patches according to the basic segmentation length. Based on financial event patches and basic financial patches, dynamic financial patch sequences are generated in the dynamic financial patch construction unit by arranging the financial event patches and basic financial patches in the time order of standard financial time series data.

[0026] In this embodiment, generating the financial embedding sequence of the account group includes: Based on the dynamic financial patch sequence, in the subject group patch embedding unit, the accounting subject balance data, transaction flow data, accounts receivable aging data, accounts payable due data, cash flow data, revenue data, expense data and liability data in each financial event patch and the basic financial patch are extracted, and the time order of each data in the standard financial time series data is preserved. Based on the unified account code and the superior account code in the accounting account balance data, the account affiliation relationship is determined, unified account codes belonging to the same superior account code are grouped into the same account group, and an account group number is assigned to each account group. Based on the debit and credit entries in the accounting subject balance data, the debit and credit direction relationship is determined. The data corresponding to the debit entries is marked as the debit direction, and the data corresponding to the credit entries is marked as the credit direction. The debit and credit directions are then written into the data records under the corresponding subject group number. Based on the payer, payee, transaction amount and transaction time in the transaction flow data, determine the fund flow relationship, arrange the fund flow order from payer to payee according to the transaction time, and write the fund flow order into the data record under the corresponding account group number; Based on the time sequence of accounts receivable aging data, accounts payable maturity data and standard financial time series data, the maturity relationship of accounts receivable is determined, and the change position of the aging level corresponding to the accounts receivable aging data and the maturity position corresponding to the accounts payable maturity data are written into the data record under the corresponding account group number. Based on the relationship between account affiliation, debit and credit direction, fund flow, and accounts receivable maturity, the data in the dynamic financial patch sequence is divided into embedding channels according to the account group number. The debit direction and credit direction are converted into embedding direction identifiers, the fund flow sequence is converted into embedding connection sequence, the change position of the aging level and the maturity position are converted into time index offset identifiers, and the embedding direction identifier, embedding connection sequence, and time index offset identifier are combined into a relationship identifier vector. Based on embedded channels and relation identifier vectors, accounting subject balance data, transaction flow data, accounts receivable aging data, accounts payable due data, cash flow data, revenue data, expense data and liability data are converted into financial data value vectors. The financial data value vectors are then combined with the corresponding relation identifier vectors to generate the subject group Patch embedded representation. Based on the subject group patch embedding representation, the subject group patch embedding units are arranged according to the time order corresponding to the subject group patch embedding representation to generate the subject group financial embedding sequence.

[0027] In this embodiment, the generation of financial risk transmission codes includes: Based on the financial embedding sequence of subject group, the subject group Patch embedding representation under the same time index is used as the attention encoding object, and the embedding sequence to be encoded is formed according to the time order corresponding to the subject group Patch embedding representation. Based on the relationship between account groups, the connection of the account group patch embedding representations in the sequence to be encoded is restricted. The account group patch embedding representations that have account affiliation relationship, lending direction relationship, fund flow relationship, and account maturity relationship are marked as connectable objects, while the account group patch embedding representations that do not have account affiliation relationship, lending direction relationship, fund flow relationship, and account maturity relationship are marked as non-connectable objects, and an attention connection mask is generated. Based on the debit and credit direction relationship and fund flow relationship in the account group association, directional constraints are set for connectable objects. The directional constraints are arranged according to the transmission direction of fund outflow to fund inflow, revenue to accounts receivable, accounts receivable to cash flow, cash flow to accounts payable, and accounts payable to liabilities, generating connectable objects with directional constraints. Based on the financial event triggering sequence, the financial event triggering intensity corresponding to the financial event triggering point is written into the connectable object with directional constraints, and the encoding weight of the connectable object with directional constraints in the triggering conduction attention encoding unit is adjusted in descending order of financial event triggering intensity to generate triggering weighted connectable objects. Based on the stimulating weighted connection objects, attention encoding is performed on the embedded sequence to be encoded in the stimulating transmission attention encoding unit. Unconnectable objects are excluded from attention connections, and connectable objects with directional constraints are aggregated according to the encoding weights to generate the transmission relationship of financial event stimulating points in the subject group relationship. Based on the transmission relationship, trigger weighted connection objects with encoding weights greater than the average encoding weights within the same time index and conforming to directional constraints are selected. These trigger weighted connection objects are then connected in sequence according to time order and directional constraints to generate a risk transmission path. Based on the transmission relationship and risk transmission path, the financial event trigger point, coding weight, directional constraint, trigger weighted connection object and time sequence in the financial event trigger sequence are combined to generate a financial risk transmission coding representation.

[0028] In this implementation method, the generation of enterprise financial management risk prediction results includes: Based on the financial risk transmission coding representation, the phased progressive forecasting unit is set up with the following stages: accounts receivable occupancy forecasting stage, cash flow pressure forecasting stage, repayment concentration forecasting stage, profit erosion forecasting stage, and comprehensive risk forecasting stage. The output results of the accounts receivable occupancy forecasting stage are input into the cash flow pressure forecasting stage, the output results of the cash flow pressure forecasting stage are input into the repayment concentration forecasting stage, the output results of the repayment concentration forecasting stage are input into the profit erosion forecasting stage, and the output results of the profit erosion forecasting stage are input into the comprehensive risk forecasting stage. Based on the financial risk transmission coding representation, the coding content corresponding to accounts receivable aging data and revenue data is extracted in the accounts receivable occupancy prediction stage. The coded content is used to predict the extended aging status of accounts receivable, the delayed collection status after revenue recognition, and the extended occupancy time status of accounts receivable, thereby generating the accounts receivable occupancy risk status. Among them, the extended aging status of accounts receivable refers to the same accounts receivable record in the accounts receivable aging data moving from a shorter aging range to a longer aging range. The delayed collection status after revenue recognition refers to the transaction flow data corresponding to the revenue that has been recognized in the revenue data not forming a collection record within the agreed collection period. The extended occupancy time status refers to the interval between the revenue recognition time index and the actual collection time index exceeding the agreed collection period of the corresponding accounts. Based on the financial risk transmission coding representation and the accounts receivable occupancy risk status, the coding content corresponding to cash flow data and transaction flow data is extracted in the cash flow pressure prediction stage. The state of reduced operating cash inflow, increased operating cash outflow, and concentrated transaction flow expenditure is predicted to generate the cash flow pressure risk status. Based on the financial risk transmission coding representation and the cash flow pressure risk status, the coding content corresponding to accounts payable maturity data and liability data is extracted in the solvency concentration forecasting stage. The solvency concentration risk status is then generated by predicting the accounts payable maturity concentration status, liability repayment concentration status and cash flow insufficient coverage status. Based on the financial risk transmission coding representation and the solvency concentration risk status, the coding content corresponding to the revenue data and expense data is extracted in the profit erosion prediction stage. The revenue decline status, expense increase status and profit margin compression status are predicted to generate the profit erosion risk status. Among them, the profit margin compression status refers to the state where the revenue data does not grow synchronously and the expense data continues to increase, or the decline in the revenue data is greater than the decline in the expense data, resulting in the continuous reduction of the difference between the revenue data and the expense data. Based on the financial risk transmission code representation, accounts receivable occupancy risk status, cash flow pressure risk status, solvency concentration risk status, and profit erosion risk status, a progressive summary is performed in the comprehensive risk prediction stage to generate a comprehensive financial management risk status. Based on the risk status of accounts receivable ties, cash flow pressure, solvency concentration, profit erosion, and comprehensive financial management, the results are combined according to the predicted values ​​corresponding to each risk status and the progressive order of the risk status to generate the enterprise financial management risk prediction results.

[0029] In this embodiment, generating financial risk warning information includes: Based on the enterprise financial management risk prediction results, the predicted values ​​corresponding to the accounts receivable occupancy risk status, cash flow pressure risk status, repayment concentration risk status, profit erosion risk status and comprehensive financial management risk status are extracted, and the risk levels are divided into low risk level, medium risk level, high risk level and severe risk level according to the predicted values ​​from low to high, thus generating risk levels. Based on the risk status with the highest predicted value in the enterprise financial management risk forecast results, the corresponding financial basic data is determined. The accounts receivable aging data and revenue data correspond to the accounts receivable aging risk status; the cash flow pressure risk status corresponds to cash flow data and transaction flow data; the payment concentration risk status corresponds to accounts payable due data and liability data; and the profit erosion risk status corresponds to revenue data and expense data. By combining the accounting account balance data, transaction flow data, accounts receivable aging data, accounts payable due data, cash flow data, revenue data, expense data, and liability data corresponding to the financial management risk status, risk source data is generated. Based on the financial event trigger sequence, financial event trigger points corresponding to risk source data are extracted, and the time index and data position of the financial event trigger points in standard financial time series data are determined as risk trigger positions. Based on the financial risk transmission coding representation, the risk transmission path corresponding to the risk triggering location is extracted, and the financial event triggering point, coding weight, directional constraint, triggering weighted connection object and time sequence in the financial event triggering sequence are determined from the risk transmission path. The extracted risk transmission path is used as the risk transmission path field in the financial risk early warning information. Based on the risk trigger location and risk transmission path fields, the risk trigger time index corresponding to the risk trigger location in the standard financial time series data, the transmission start time index corresponding to the first triggered weighted connection object in the risk transmission path field, and the transmission end time index corresponding to the last triggered weighted connection object in the risk transmission path field are extracted. The earliest time index among the risk trigger time index, transmission start time index, and transmission end time index is determined as the risk occurrence start time index, and the latest time index among the risk trigger time index, transmission start time index, and transmission end time index is determined as the risk occurrence end time index. The continuous time range between the risk occurrence start time index and the risk occurrence end time index is determined as the risk occurrence time range. Based on the risk level, risk source data, risk trigger location, risk transmission path field, and risk occurrence time range, the fields are combined according to the same risk status. The risk level is written into the warning level field, the risk source data is written into the warning source field, the risk trigger location is written into the warning trigger field, the risk transmission path field is written into the warning transmission field, and the risk occurrence time range is written into the warning time range field to generate financial risk warning information.

[0030] Example 1: To verify the feasibility of this invention in practice, it was applied to a financial management risk prediction scenario for a group company. This company has multiple business segments and numerous sources of daily financial data, including accounting balance data, transaction flow data, accounts receivable aging data, accounts payable due data, cash flow data, revenue data, expense data, and liability data. Due to inconsistencies in the statistical standards of different business systems, some transaction flows are delayed in being recorded, some accounts receivable are not collected promptly after revenue recognition, and some accounts payable have concentrated due dates. Traditional financial management methods mainly rely on manual report verification and fixed threshold warnings, which easily leads to problems such as delayed risk detection, unclear risk source identification, and inability to identify the transmission relationship between accounts receivable ties and cash flow pressure risks.

[0031] In this embodiment, 18 consecutive months of financial data were selected as verification data, totaling 486,000 original records. These included 43,200 accounting balance records, 268,000 transaction records, 52,600 accounts receivable aging records, 48,700 accounts payable due records, 28,100 cash flow records, 23,600 revenue records, 15,800 expense records, and 6,000 liability records. After performing field cleaning, time alignment, account mapping, and missing value imputation on the above data, 4,280 duplicate transaction records were deleted, 920 abnormal date records were corrected, 3,180 non-critical missing fields were filled, and 1,460 records lacking critical fields were removed, resulting in 479,340 valid records and generating standard financial time-series data. This standard financial time-series data is organized according to a unified time standard, enabling comparison of revenue recognition, cash collection, changes in accounts receivable aging, changes in cash flow gaps, and accounts payable due dates on the same timeline.

[0032] In practical applications, standard financial time-series data is input into the trained PatchTST model. The financial event triggering unit identifies financial event trigger points based on changes in account balances, concentrated transaction flows, aging migration, accounts payable maturity, and widening cash flow gaps. After processing, a total of 1376 financial event trigger points are identified, including 312 changes in account balances, 286 concentrated transaction flows, 268 aging migrations, 241 accounts payable maturity, and 269 widening cash flow gaps. The dynamic financial patch construction unit generates a dynamic financial patch sequence around these trigger points, resulting in 1376 financial event patches and 772 basic financial patches. Compared to fixed-length segmentation, this approach reduces the likelihood of financial events being split into different time segments, ensuring that aging migration, widening cash flow gaps, and concentrated accounts payable maturity remain fully represented within continuous segments.

[0033] In the account group financial embedding stage, the account group patch embedding unit organizes a dynamic financial patch sequence according to account affiliation, debit / credit direction, cash flow, and accounts receivable maturity, converting scattered financial data into an account group patch embedded representation. The attention-based encoding unit further encodes the transmission relationship of financial event triggers within the account group relationships, obtaining a financial risk transmission encoding representation. Verification showed that 31.6% of the samples experienced accounts receivable pressure risk due to delayed revenue recognition; 26.8% experienced cash flow pressure risk due to concentrated transaction flows and widening cash flow gaps; 21.4% experienced payment concentration risk due to concentrated accounts payable maturities; and 20.2% experienced profit erosion risk due to declining revenue and rising expenses. These results indicate that corporate financial management risk is not solely caused by a single abnormal financial indicator, but rather by the continuous transmission of multiple financial events along account group relationships.

[0034] In the risk prediction phase, the progressive prediction unit, based on the financial risk transmission code, sequentially generates the following risk states: accounts receivable pressure risk, cash flow pressure risk, concentrated repayment risk, profit erosion risk, and comprehensive financial management risk. During the verification period, a total of 927 financial risk warning messages were generated, including 126 high-risk warning messages, 284 medium-risk warning messages, and 517 low-risk warning messages. After financial review, 109 of the high-risk warning messages were consistent with subsequent accounts receivable overdue, cash flow gaps, concentrated repayment pressure, and abnormal profit situations; 232 of the medium-risk warning messages were consistent with subsequent financial fluctuations. These results indicate that the method of this invention can provide relatively stable warning information before risks occur, and simultaneously outputs the risk level, risk source data, risk trigger location, risk transmission path, and risk occurrence time range.

[0035] To verify the practical effectiveness of the method of this invention, this embodiment sets up three methods for comparison. Method 1 is the traditional fixed threshold statistical method, which judges risk through financial indicator thresholds, aging thresholds, and cash flow gap thresholds; Method 2 is the traditional fixed patch time series prediction method, which performs time series prediction after dividing financial time series data into fixed time lengths; Method 3 is the method of this invention, namely a corporate financial management risk prediction method based on big data analysis. All three methods use the same batch of standard financial time series data, with training samples accounting for 70%, validation samples accounting for 15%, and test samples accounting for 15%. The evaluation results are shown in Table 1.

[0036] Table 1 Comparison of Enterprise Financial Management Risk Prediction Effects

[0037] As shown in Table 1, the financial event identification accuracy of the method of this invention is 91.2%, which is 9.5 percentage points higher than the traditional threshold rule method and 4.8 percentage points higher than the fixed window PatchTST method. The performance improvement is due to the fact that the method of this invention does not mechanically segment financial data according to a fixed time length, but forms financial event patches around the triggering point of the financial event. This allows sudden changes in accounting balances, aging migration, accounts payable due, and widening cash flow gaps to be continuously expressed in the same risk segment, reducing the information loss caused by the truncation of financial events by the time window.

[0038] The risk prediction accuracy of the method in this invention is 89.1%, and the high-risk warning hit rate is 87.4%, both higher than the other two comparative methods. Traditional threshold rule methods mainly rely on the judgment of a single indicator exceeding the limit, which cannot express the transmission process between revenue recognition, delayed collection, widening cash flow gap, and concentrated accounts payable maturity; although the fixed window PatchTST method can handle time series changes, its patch division method does not pay attention to the trigger position of financial events, and easily disperses key risk changes across different windows. The method in this invention combines the trigger point of financial events with the relationship between account groups through the embedded representation of account group patches and the encoding representation of financial risk transmission, which can more accurately identify the process of risk transmission from accounts receivable to cash flow pressure, concentrated repayment, and profit erosion.

[0039] The average advance warning period of the method in this invention is 8.1 days, which is 2.9 days longer than the traditional threshold rule method and 1.3 days longer than the fixed-window PatchTST method. The false alarm rate is 11.6%, lower than both the traditional threshold rule method and the fixed-window PatchTST method. This result indicates that the method in this invention does not simply increase the number of warnings, but rather detects risk changes in advance while maintaining a low false alarm rate. This is because the progressive prediction unit predicts the accounts receivable pressure risk, cash flow pressure risk, solvency concentration risk, profit erosion risk, and comprehensive financial management risk in a progressive relationship. The preceding risk status can participate in the judgment of the following risk status, avoiding the direct attribution of short-term financial fluctuations to comprehensive financial management risk.

[0040] The batch processing time of the method in this invention is 31.2 minutes, slightly longer than the 28.4 minutes of the fixed-window PatchTST method, but significantly lower than the 46.5 minutes of the traditional threshold rule method. This time difference is within a reasonable range because the method in this invention adds financial event trigger point identification, dynamic financial patch sequence construction, and financial risk transmission encoding processing. The calculation process is more complex than the fixed-window PatchTST method, but it achieves higher prediction accuracy, higher high-risk warning hit rate, and lower false alarm rate. For enterprise financial management risk prediction scenarios, this processing time can meet the needs of daily batch analysis and periodic risk warnings.

[0041] The above are merely preferred embodiments 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 enterprise financial management risks based on big data analysis, characterized in that, Includes the following steps: Collect basic financial data, perform field cleaning, time alignment, account mapping and missing value filling to generate standard financial time series data; Construct a financial PatchTST model, including a financial event triggering unit, a dynamic financial Patch construction unit, a subject group Patch embedding unit, a triggering and transmission attention encoding unit, and a stage-progressive prediction unit; Based on standard financial time-series data, financial event triggering points are identified in the financial event triggering unit, and a financial event triggering sequence is generated. Based on the financial event triggering sequence, standard financial time series data is converted into financial event patches around the financial event triggering point through dynamic financial patch construction units, generating a dynamic financial patch sequence. Based on the dynamic financial patch sequence, the relationship between account groups is constructed through the account group patch embedding unit, and the dynamic financial patch sequence is organized into an account group patch embedding representation to generate the account group financial embedding sequence. Based on the financial embedding sequence of subject groups, the transmission relationship of financial event triggering points in the subject group relationship is encoded by stimulating the transmission attention encoding unit, thereby generating a financial risk transmission encoding representation; Based on the financial risk transmission coding representation, risk prediction is performed through a phased progressive prediction unit to generate enterprise financial management risk prediction results; Based on the enterprise financial management risk prediction results, financial event triggering sequence and financial risk transmission code representation, financial risk early warning information is generated.

2. The enterprise financial management risk prediction method based on big data analysis according to claim 1, characterized in that, The generated standard financial time-series data includes: Collect basic financial data, including accounting balance data, transaction flow data, accounts receivable aging data, accounts payable due data, cash flow data, revenue data, expense data, and liability data, and write them into the same data processing space according to the data source identifier to form the basic financial data to be cleaned; Based on the financial basic data to be cleaned, perform field cleaning to generate cleaned financial basic data; Based on the cleaned financial data, time alignment is performed, and data is written to the same time position to generate time-aligned financial data. Based on time-aligned financial basic data, perform account mapping, write it under the corresponding unified account code, and generate account mapping financial basic data; Based on the financial basic data of account mapping, identify the location of missing data, perform missing value filling, and generate the filled financial basic data; Based on the supplemented financial data, fields are combined to generate standard financial time-series data.

3. The enterprise financial management risk prediction method based on big data analysis according to claim 1, characterized in that, The construction of the financial PatchTST model includes: The basic model structure is built based on the PatchTST model, including the Patch partitioning structure, Patch embedding structure, positional encoding structure, Transformer encoding structure, and prediction head structure. Based on the basic model structure, a financial event triggering unit is added; Based on the Patch partitioning structure, the partitioning method of forming Patch input markers according to a fixed time length is replaced by a dynamic financial Patch construction method that adjusts the boundaries of continuous time segments according to the financial event triggering sequence, and a dynamic financial Patch construction unit is constructed. Based on the Patch embedding structure, the processing method of Patch embedding on a single data channel is replaced by the processing method of Patch embedding on dynamic financial Patch sequences based on the relationship of account groups, and account group Patch embedding units are constructed. Based on the Transformer encoding structure, the self-attention connection structure is replaced with an attention connection structure that introduces the relationship between the financial event triggering sequence and the subject group, and a triggering and conduction attention encoding unit is constructed. Based on the forecast head structure, the single output structure that directly outputs the time series forecast results is replaced with a forecast structure that outputs progressively according to the status of accounts receivable occupancy risk, cash flow pressure risk, repayment concentration risk, profit erosion risk and comprehensive financial management risk, thus constructing a stage-progressive forecast unit. Based on the connection relationship between the financial event triggering unit, the dynamic financial patch construction unit, the subject group patch embedding unit, the triggering transmission attention encoding unit, and the stage-progressive prediction unit, the triggering financial patchTST model is generated.

4. The enterprise financial management risk prediction method based on big data analysis according to claim 1, characterized in that, The generated financial event trigger sequence includes: Based on standard financial time-series data, in the financial event triggering unit, based on the accounting account balance data, adjacent data values ​​are compared according to the time index to determine the position of the accounting account balance change. Based on transaction flow data, the number of transactions, total transaction amount, and concentration of transaction objects are statistically analyzed to determine the location of concentrated transaction flow. Based on accounts receivable aging data, the aging levels are compared according to the time index to determine the aging migration position; Based on accounts payable due date data, the due date is matched to determine the due date position of accounts payable; Based on cash flow data, extract the values ​​of operating cash inflows and operating cash outflows to determine the cash flow gap and identify the location where the cash flow gap is widening. Based on the location of changes in accounting account balances, the location of concentrated transaction flows, the location of aging migration, the location of accounts payable due, and the location of widening cash flow gaps, the location is merged in the financial event triggering unit. The time index after location merging is determined as the financial event triggering point, and the corresponding financial event triggering intensity is generated. Based on the trigger points and intensity of financial events, financial event trigger sequences are generated by arranging the trigger points in the time order of standard financial time series data.

5. The enterprise financial management risk prediction method based on big data analysis according to claim 1, characterized in that, The generation of the dynamic financial patch sequence includes: Based on standard financial time-series data, in the dynamic financial patch construction unit, continuous time segments are divided according to the time index sequence and the basic segmentation length, and the merging interval threshold and the number of overlapping time indices are determined. Based on the financial event trigger sequence, the time index corresponding to the financial event trigger point is matched with the time index range corresponding to the continuous time segment to generate the continuous time segment matching result. Based on the continuous time segment matching results, select the continuous time segment containing the financial event trigger point, and determine the time index corresponding to the financial event trigger point as the financial event patch center; Based on the financial event patch center, the boundaries of continuous time segments are adjusted and merged into financial event patches; Based on the continuous time segment matching results, continuous time segments that do not contain financial event trigger points are selected and converted into basic financial patches. A dynamic financial patch sequence is generated based on financial event patches and basic financial patches.

6. The enterprise financial management risk prediction method based on big data analysis according to claim 1, characterized in that, The generated subject group financial embedding sequence includes: Based on the dynamic financial patch sequence, in the subject group patch embedding unit, the accounting subject balance data, transaction flow data, accounts receivable aging data, accounts payable due data, cash flow data, revenue data, expense data and liability data of each financial event patch and the basic financial patch are extracted. Based on the unified account code and the superior account code in the accounting account balance data, the account affiliation relationship is determined, account group numbers are assigned, and the debit and credit direction relationship is determined based on the debit and credit amounts. Based on the payer, payee, transaction amount and transaction time in the transaction log data, determine the relationship of fund flow; Based on the time sequence of accounts receivable aging data, accounts payable due data, and standard financial time series data, the due date relationships are determined. Based on the relationships of account affiliation, debit and credit direction, fund flow, and accounts receivable maturity, the data in the dynamic financial patch sequence is divided into embedded channels according to the account group number, and a relationship identifier vector is generated. Based on the embedded channels and relational identifier vectors, accounting account balance data, transaction flow data, accounts receivable aging data, accounts payable due data, cash flow data, revenue data, expense data, and liability data are converted into financial data value vectors, and then combined with the corresponding relational identifier vectors to generate the account group Patch embedded representation, which is then arranged to generate the account group financial embedded sequence.

7. The enterprise financial management risk prediction method based on big data analysis according to claim 1, characterized in that, The generated financial risk transmission code includes: Based on the financial embedding sequence of subject group, the subject group Patch embedding representation under the same time index is used as the attention encoding object to form the embedding sequence to be encoded. Based on the subject group association relationship, the subject group Patch embedding representation in the embedding sequence to be encoded is connected and constrained, and connectable and non-connectable objects are marked to generate an attention connection mask. Based on the borrowing and lending direction relationship and fund flow relationship in the subject group association relationship, directional constraints are set on connectable objects to generate connectable objects with directional constraints; Based on the financial event triggering sequence, the financial event triggering intensity is written into a connectable object with directional constraints, the encoding weights are adjusted, and a triggering weighted connectable object is generated. Based on the excitation weighted connection object, attention encoding is performed on the embedded sequence to be encoded in the excitation transmission attention encoding unit to generate the transmission relationship of financial event excitation points in the account group relationship; Based on the transmission relationship, generate risk transmission paths; Based on the transmission relationship and risk transmission path, the financial event trigger point, coding weight, directional constraint, trigger weighted connection object and time sequence in the financial event trigger sequence are combined to generate a financial risk transmission coding representation.

8. The enterprise financial management risk prediction method based on big data analysis according to claim 1, characterized in that, The generated enterprise financial management risk prediction results include: Based on the financial risk transmission coding representation, the phased progressive forecasting unit is set up with the following stages: accounts receivable occupancy forecasting stage, cash flow pressure forecasting stage, repayment concentration forecasting stage, profit erosion forecasting stage, and comprehensive risk forecasting stage. Based on the financial risk transmission coding representation, the coding content corresponding to the accounts receivable aging data and revenue data is extracted in the accounts receivable occupancy prediction stage to generate the accounts receivable occupancy risk status. Based on the financial risk transmission coding representation and the accounts receivable occupancy risk status, the coding content corresponding to cash flow data and transaction flow data is extracted in the cash flow pressure forecasting stage to generate the cash flow pressure risk status. Based on the financial risk transmission coding representation and the cash flow pressure risk status, the coding content corresponding to the accounts payable maturity data and liability data is extracted in the solvency concentration forecasting stage to generate the solvency concentration risk status. Based on the financial risk transmission coding representation and the solvency concentration risk status, the coding content corresponding to the revenue and expense data is extracted in the profit erosion prediction stage to generate the profit erosion risk status. Based on the financial risk transmission code representation, accounts receivable occupancy risk status, cash flow pressure risk status, solvency concentration risk status, and profit erosion risk status, a progressive summary is performed in the comprehensive risk prediction stage to generate a comprehensive financial management risk status. Based on the risk status of accounts receivable ties, cash flow pressure, payment concentration, profit erosion, and comprehensive financial management risk, a combination of risk status is used to generate a prediction result of corporate financial management risk.

9. The enterprise financial management risk prediction method based on big data analysis according to claim 1, characterized in that, The generated financial risk warning information includes: Based on the enterprise financial management risk prediction results, the predicted values ​​corresponding to the status are extracted and divided into low risk level, medium risk level, high risk level and severe risk level to generate risk level; Based on the risk status with the highest predicted value in the enterprise financial management risk prediction results, determine the financial basic data corresponding to the risk status and generate risk source data; Based on the financial event trigger sequence, extract the financial event trigger points corresponding to the risk source data and determine them as risk trigger locations; Based on the financial risk transmission code representation, the risk transmission path corresponding to the risk triggering location is extracted and used as the risk transmission path field in the financial risk early warning information; Based on the risk trigger location and risk transmission path fields, extract the risk occurrence start time index and risk occurrence end time index to determine the risk occurrence time range; Based on risk level, risk source data, risk trigger location, risk transmission path fields, and risk occurrence time range, financial risk warning information is generated by combining fields according to the same risk status.

Images