A financial budget control method and system based on big data analysis

Abstract

The application relates to the technical field of financial budget management, and discloses a financial budget control method and system based on big data analysis, wherein the method comprises the following steps: obtaining expenditure execution data flow and internal transaction flow data flow of each subsidiary company, generating a directed conduction atlas through Granger causality test; calculating a deviation trigger ratio when new data is generated at a high-frequency node, and activating correlation analysis if the deviation trigger ratio exceeds a threshold value; calculating a posteriori estimation value and an uncertainty parameter for a low-frequency reporting downstream node; inputting a node feature vector into a graph attention network, introducing an uncertainty attenuation factor to perform message passing aggregation, and outputting a cascading impact prediction value; generating a budget control signal according to the difference and the confidence interval relationship; and locally updating the prediction and outputting a final execution instruction after actual data of the low-frequency node arrives.

Description

Technical Field

[0001] This invention relates to the field of financial budget management technology, specifically to a financial budget control method and system based on big data analysis. Background Technology

[0002] Group companies typically operate through an internal supply chain, with headquarters and multiple subsidiaries forming business relationships. Upstream subsidiaries are responsible for raw material procurement, midstream subsidiaries for manufacturing, and downstream subsidiaries for product sales. Each subsidiary submits budget execution data to the group at different frequencies, with reporting cycles varying from daily to weekly to monthly. When an upstream subsidiary experiences a budget deviation, this deviation propagates downstream along the internal transaction chain, creating a cascading effect.

[0003] Existing batch big data analytics methods require waiting for data from all subsidiaries to align before performing correlation analysis. This results in a delay of several weeks in early warning signals for cascading deviations. During this waiting period, the analysis process uses historical data from low-frequency reporting nodes to fill in the gaps. Deviations between historical data and the actual situation can overestimate or underestimate the strength of the transmission effect, leading to inaccurate correlation analysis results. Therefore, the existing method faces two problems simultaneously: insufficient timeliness of budget control and decreased accuracy of correlation analysis, making it difficult to meet the precise budget control needs of large group enterprises for cascading deviations in their internal supply chains. Summary of the Invention

[0004] This invention provides a financial budget control method and system based on big data analysis, which solves the technical problems in related technologies such as the inability to timely identify the transmission effect between budget items within a group, the inaccuracy of cascading impact prediction due to missing data from low-frequency reporting nodes, and insufficient confidence in budget control signals.

[0005] This invention discloses a financial budget control method based on big data analysis, comprising: Obtain expenditure execution data streams and internal transaction flow data streams for each subsidiary's budget items. Perform Granger causality tests on the expenditure time sequence of each subsidiary's budget items based on the internal transaction flow data to generate a directed transmission graph between budget items. When the ratio of the expenditure deviation change rate of the high-frequency reporting node to the historical volatility exceeds a preset multiple threshold, it is marked as a cascading trigger event. For low-frequency downstream nodes in the directed transmission spectrum that have not yet received new data, the posterior estimate and uncertainty parameter are calculated based on the conditional probability distribution of the transmission relationship and the observed deviation sequence of the upstream node. The actual observed deviation or posterior estimated value of each node is used as the node feature vector input to the graph attention network. During the message aggregation stage, the attention weight of the posterior estimated node is attenuated and adjusted according to the uncertainty parameter, and the cascade impact prediction value and prediction confidence interval of each downstream node are output. The cascading shock forecasts are superimposed on the baseline expenditure forecasts, and a budget control signal is generated based on the difference between the cumulative expenditure forecasts and the remaining budget amount, as well as the forecast confidence interval.

[0006] Furthermore, in the directed transmission graph, nodes represent the budget items of each subsidiary, directed edges represent the transmission direction, edge weights are the statistic values ​​of the Granger causality test as the causal strength, and edge attributes include a transmission delay parameter, which is the time interval corresponding to the optimal lag order in the Granger causality test.

[0007] Furthermore, the cascading trigger event includes: calculating the difference between the actual expenditure value and the budgeted planned value of the current period of the high-frequency reporting node to obtain the current expenditure deviation value, and calculating the difference between the current expenditure deviation value and the expenditure deviation value of the previous period to obtain the expenditure deviation change rate. ; Obtain the standard deviation of the expenditure deviation sequence within the historical period of this node as the historical volatility. ; Calculate the deviation trigger ratio ,when The time marker is designated as a cascading event; where, For nodes The rate of change in expenditure deviation, For nodes Historical volatility, For the preset multiple threshold, Number the nodes.

[0008] Further, the calculation of the posterior estimate and uncertainty parameter includes: traversing along the edge direction of the directed transmission graph from the node that triggered the cascading event, marking nodes that have not reported new data within the current analysis window as nodes to be inferred; for each node to be inferred... Get the set of all its updated upstream parent nodes. Extract each upstream parent node In conduction delay Previous observed deviation values According to the strength of causation Weighted posterior estimate: ;in, For the current analysis time, For nodes To node The propagation delay parameter, For nodes To node Causal strength; uncertainty parameter Based on the deviations of each upstream parent node in historical data and the node to be inferred The residual variance is obtained by calculating the conditional distribution between deviations.

[0009] Furthermore, when the upstream parent node of the node to be inferred is also the node to be inferred, the posterior estimate of the upstream parent node is used to replace the actual observation value in the calculation, and the uncertainty parameter of the upstream parent node is added to the uncertainty parameter of the current node to be inferred; when the time interval between the latest data of the low-frequency reported downstream node and the current time exceeds a preset multiple of the reporting period of the node, the uncertainty parameter of the node is set to the upper limit value.

[0010] Furthermore, the attenuation adjustment of the attention weights for the posterior estimation nodes according to the uncertainty parameter includes: for nodes originating from the node to be inferred... The message's attention weight is multiplied by a decay factor. ,in For the message source node Uncertainty parameters after normalization; when When the attenuation factor is zero, it is 1. As the factor increases, the attenuation factor approaches zero.

[0011] Furthermore, the prediction confidence interval is Among them, the boundary value of prediction uncertainty The calculation method is as follows: ,in Pointing to a node The set of all upstream message paths, For one of the paths, For directed edges on the path The causal strength, For nodes on the path The uncertainty parameter after normalization The total number of paths, For nodes The predicted value of cascading impacts.

[0012] Furthermore, the generation of the budget control signal includes: calculating the cumulative expenditure forecast. With remaining budget The difference ;when Exceeding the budget control threshold And the lower bound of the prediction confidence interval satisfies At that time, a high-confidence budget control signal is generated; when Exceeding the budget control threshold However, the lower bound of the prediction confidence interval is satisfied. At that time, a budget warning signal awaiting confirmation is generated; among which For nodes Baseline expenditure forecasts, The predicted values ​​of the cascade impacts are after inverse normalization. To predict the uncertainty boundary values, Number the downstream node.

[0013] Furthermore, it also includes: when the actual data of the low-frequency reporting node arrives, replacing the posterior estimate of the node with the actual observation and updating the uncertainty parameter to zero; determining the affected local subgraph range with the data update node as the center; re-performing the message passing aggregation operation of the graph attention network only for the nodes within the local subgraph; updating the cascaded impact prediction value and prediction confidence interval; correcting or confirming the generated budget control signal; and outputting the final budget control execution instruction.

[0014] This invention discloses a financial budget control system based on big data analysis, used to execute the aforementioned financial budget control method based on big data analysis, comprising: The data acquisition and graph generation module is used to acquire expenditure execution data flow and internal transaction flow data flow of budget items of various subsidiaries of the group, perform Granger causality test based on internal transaction flow data, and generate directed transmission graph between budget items; The cascading trigger detection module is used to calculate the ratio of the expenditure deviation change rate to the historical volatility when a new data event is generated at the high-frequency reporting node. When the ratio exceeds a preset multiple threshold, it is marked as a cascading trigger event. The low-frequency node state inference module is used to calculate the posterior estimate and uncertainty parameters for low-frequency downstream nodes in the directed transmission spectrum that have not yet received new data, based on the conditional probability distribution of the transmission relationship and the observed deviation sequence of the upstream node. The cascaded impact prediction module is used to input the actual observed deviation value or posterior estimated value of each node as the node feature vector into the graph attention network. During the message aggregation stage, the attention weight of the posterior estimated node is attenuated and adjusted according to the uncertainty parameter, and the cascaded impact prediction value and prediction confidence interval of each downstream node are output. The budget control signal generation module is used to superimpose the cascading shock forecast value onto the baseline expenditure forecast value, and generate a budget control signal based on the difference between the cumulative expenditure forecast value and the remaining budget amount and the forecast confidence interval.

[0015] The present invention has the following beneficial effects: By extracting the directed transmission graph between budget items from internal transaction flow data through Granger causality test, big data correlation analysis can trace the transmission path of cascading deviations along the directional causal relationship of the internal supply chain, overcoming the limitation that the budget data of each subsidiary is processed in isolation and that the cascading effect across subsidiaries cannot be perceived.

[0016] By performing posterior state inference on the downstream nodes of low-frequency reports, the graph attention network can perform cascade propagation analysis without waiting for all node data to align. This advances the response time of the budget control signal from when the low-frequency node data arrives completely to when the high-frequency data event occurs, eliminating the analysis delay caused by the difference in reporting frequency.

[0017] By introducing an uncertainty decay factor in the message aggregation stage of the graph attention network, the message contribution of the posterior estimated node is automatically adjusted according to the credibility of its inference, avoiding the overestimation or underestimation of the transmission effect by historical data filling, and ensuring that the accuracy of budget control decisions is not significantly reduced due to the lack of data from some nodes. Attached Figure Description

[0018] Figure 1 This is a flowchart of a financial budget control method based on big data analysis provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the node relationship diagram of the directed transmission spectrum provided in an embodiment of the present invention; Figure 3 This is a schematic diagram comparing the trigger detection indicators of Company A's procurement items provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the posterior estimate and uncertainty parameter of low-frequency nodes provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the attenuation factor of each node and the predicted value of the cascaded impact in the graph attention network provided in this embodiment of the invention; Figure 6 This is a schematic diagram illustrating the comparison of the budget execution status of downstream nodes provided in an embodiment of the present invention; Figure 7 This is a schematic diagram showing the comparison between the budget control signal difference and the threshold provided in an embodiment of the present invention; Figure 8 This is a schematic diagram illustrating the change in the boundary value of prediction uncertainty before and after data update, provided in an embodiment of the present invention. Figure 9 This is a schematic diagram of the data reporting frequency and response delay distribution of the three-tier supply chain subsidiaries provided in this embodiment of the invention. Detailed Implementation

[0019] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0020] At least one embodiment of the present invention discloses a financial budget control method based on big data analysis, see [link to relevant documentation]. Figure 1 This includes the following steps: Step 1: Obtain budget execution data and generate a directed transmission graph: Acquire expenditure execution data streams and internal transaction log data streams for each subsidiary's budget items. Register update frequency tags and latest arrival timestamps for each data stream. Based on the internal transaction log data, perform Granger causality tests on the expenditure time series of each subsidiary's budget items. This Granger causality test examines the directional causal relationship between the expenditure time series. Based on the Granger causality test results, generate a directed transmission graph among the group's budget items. In this directed transmission graph, nodes represent each subsidiary's budget item, directed edges represent the transmission direction, edge weights are causal strength values, and edge attributes include transmission delay parameters.

[0021] Step 1 may specifically include: Step S101: Collect expenditure execution data streams from each subsidiary. Divide the expenditure execution data streams into high-frequency and low-frequency data streams according to the reporting cycle. Record an update frequency tag and the latest arrival timestamp for each expenditure execution data stream. The update frequency tag includes three categories: daily, weekly, and monthly.

[0022] Step S102: Collect the group's internal transaction flow data stream. Extract transaction relationship records between subsidiaries. Obtain the internal transaction amount sequence between subsidiaries arranged by time.

[0023] Step S103: Based on internal transaction data, use the Granger causality test to pairwise examine the expenditure time series of each subsidiary's budget items. For each pair of expenditure time series, determine whether the historical expenditure data of one item has a significant predictive power for the current expenditure of the other item. When the significance level of the Granger causality test statistic is lower than a preset threshold, a directional causal relationship is confirmed between the two, and the corresponding significance level and optimal lag order are recorded.

[0024] Step S104: Connect the pairs of items that pass the Granger causality test as directed edges. Summarize all directed edges to generate a directed transmission graph. The weight of each directed edge is taken as the causal strength from the statistic value in the Granger causality test. The transmission delay parameter is recorded in the attributes of each directed edge. This transmission delay parameter is taken as the time interval corresponding to the optimal lag order in the Granger causality test.

[0025] It should be noted that the aforementioned transmission delay parameter refers to the time length corresponding to the lag order that minimizes the prediction error of downstream node expenditures after applying different lag orders to the upstream node expenditure time series in the Granger causality test and performing regression analysis. This time length reflects the time required for upstream budget deviations to be transmitted to downstream nodes. For example, if the upstream subsidiary's procurement item... The deviation of the week affects the manufacturing item of the midstream subsidiary in the first week. If weekly expenditures have the most significant predictive power, then the propagation delay parameter is two weeks. Among these, This is the sequence number of the time period.

[0026] Step 2, detect cascading trigger events: When a new data event is generated in the expenditure execution data stream of any high-frequency reporting node, the ratio of the expenditure deviation change rate of that high-frequency reporting node to its historical volatility is calculated to obtain the deviation trigger ratio. When the deviation trigger ratio exceeds a preset multiple threshold, the event is marked as a cascading trigger event, activating the big data correlation analysis process.

[0027] Step 2 may specifically include: Step S201: Receive newly arrived expenditure execution data from the high-frequency reporting node. Calculate the difference between the actual expenditure value and the budgeted value for the current period of the high-frequency reporting node to obtain the current expenditure deviation value. Calculate the difference between the current expenditure deviation value and the expenditure deviation value of the previous period to obtain the expenditure deviation change rate. .in, The unit of measurement should be consistent with the unit of expenditure amount, which is a monetary unit. Number the nodes.

[0028] Step S202: Obtain the standard deviation of the expenditure deviation sequence within the historical period of the high-frequency reporting node, as the historical volatility. . Units and Both are in monetary units and have the same dimensions, so they can be directly used for ratio calculations. (Calculation of deviation trigger ratio) : ; in, For nodes The rate of change in expenditure deviation, For nodes Historical volatility, It is a dimensionless ratio. Number the nodes.

[0029] Step S203: Adjust the deviation trigger ratio Compared with the preset multiple threshold Comparison. When When this occurs, the event is marked as a cascading trigger event, activating the subsequent correlation analysis process. At this time, no correlation analysis is triggered; only the historical deviation record of the high-frequency reporting node is updated.

[0030] It should be noted that the above-mentioned preset multiple threshold It can be set based on the statistical patterns of cascading deviation events in the group's historical data. For example, it can be... Setting it to 2 means that an abnormal change is considered to occur when the rate of change in expenditure deviation exceeds twice the historical volatility. Different settings can also be applied to subsidiaries with different business types. value.

[0031] Step 3, infer the current deviation state of the low-frequency node: For the activated association analysis process, low-frequency reporting downstream nodes that have not yet received new data are identified in the directed transmission graph. Using the conditional probability distribution of the transmission relationship between the low-frequency reporting downstream node and the updated upstream node, combined with the transmission delay parameter and the observed deviation sequence of the upstream node, the posterior estimate and uncertainty parameter of the expenditure deviation of the low-frequency reporting downstream node at the current time are calculated.

[0032] Step 3 may specifically include: Step S301: Traverse the directed transmission graph along the edge directions, starting from the high-frequency reporting node that triggered the cascading event. The traversal range includes direct downstream nodes and indirect downstream nodes. Check the latest arrival timestamp of each downstream node. Mark nodes that have not reported new data within the current analysis window as nodes to be inferred.

[0033] Step S302: For each node to be inferred Get the set of all updated upstream parent nodes. Extract each upstream parent node. In conduction delay The previously observed bias sequence. Simultaneously, the causal strength of the corresponding directed edges is obtained. .in, Let be the node number to be inferred. This is the number of the upstream parent node. For nodes To node The propagation delay parameter, For nodes To node Causality strength of directed edges.

[0034] Step S303: Calculate the node to be inferred based on the conditional probability distribution of the transmission relationship. At the present moment posterior estimate and uncertainty parameters .

[0035] Posterior estimate Observed deviation values ​​from each upstream parent node Having the same monetary units, causal strength These are dimensionless weighting coefficients. The weighted sum is divided by the sum of the weights to maintain consistent dimensions. Posterior estimate. The calculation method is as follows: Each upstream parent node... At any moment Observed deviation value According to the corresponding causal strength Perform a weighted summation, then divide by the sum of the causal strengths of all upstream parent nodes involved in the calculation, specifically: ; in, For nodes The set of all updated upstream parent nodes, For nodes At any moment The observed deviation value, For nodes To node The causal strength, For nodes To node The propagation delay parameter, For the current analysis time, Let be the node number to be inferred. This is the index of the upstream parent node. Summation and traversal of the set. All upstream parent nodes.

[0036] Uncertainty parameter Based on the deviations of each upstream parent node in historical data and the node to be inferred The residual variance is obtained by calculating the conditional distribution of deviations. Specifically, it is calculated using the deviation values ​​of each upstream parent node in the historical data, weighted by causal strength, as the independent variable and the node to be inferred. The actual deviation is used as the dependent variable for linear regression, and the sample variance of the regression residuals is taken as the uncertainty parameter. This reflects the average deviation between the posterior estimate and the true value. Uncertainty parameter. The unit is the square of the monetary unit, which corresponds to the dimension of the deviation value.

[0037] It should be noted that the above conditional probability distribution is obtained by fitting the joint distribution of upstream node deviation and downstream node deviation in historical data. When the node to be inferred has multiple upstream parent nodes, the transmission effects of each upstream parent node are weighted and superimposed according to causal strength. When the upstream parent node is also the node to be inferred, the posterior estimate of the upstream parent node is used to replace the actual observation value in the calculation, and the uncertainty parameter of the upstream parent node is added to the uncertainty parameter of the current node to be inferred.

[0038] In this embodiment, to avoid inaccurate inferences due to the expiration of the propagation delay parameter, when the time interval between the latest data of the low-frequency reporting downstream node and the current time exceeds a preset multiple of the reporting period of the low-frequency reporting downstream node, the uncertainty parameter of the low-frequency reporting downstream node is set to an upper limit value. This process ensures that the aggregation weight of the low-frequency reporting downstream node automatically decays to a minimum in subsequent steps, preventing expired inference values ​​from misleading cascaded predictions.

[0039] Step 4: Predict the cascading impact values ​​of each downstream node: The actual observed bias values ​​or posterior estimates of each node are assembled into node feature vectors, which, along with uncertainty parameters, are input into the graph attention network. The graph attention network takes the adjacency matrix of the directed transmission graph and the feature vectors of each node as input, performs message passing aggregation operations along the directed edges, and attenuates the attention weights of the posterior estimated nodes according to their uncertainty parameters during the message aggregation stage. It outputs the cascaded impact prediction values ​​and prediction confidence intervals for each downstream node.

[0040] Step 4 may specifically include: Step S401: For each node in the directed transmission graph, assemble its expenditure bias value into a node feature vector. For nodes that have reported new data, the node feature vector takes the actual observed bias value. For nodes to be inferred, the node feature vector takes the posterior estimate. At the same time, an uncertainty flag is attached to each node: the uncertainty parameter of the observed nodes is set to zero, and the uncertainty flag of the nodes to be inferred is the uncertainty parameter. Before being input into the graph attention network, the deviation values ​​of each node are normalized using Z-score standardization. This eliminates the impact of differences in expenditure scale among subsidiaries on graph propagation and aggregation operations, allowing deviation values ​​of different magnitudes to participate in message passing at the same scale.

[0041] Step S402: Input the node feature vectors and the adjacency matrix of the directed transmission graph into the graph attention network. The graph attention network is a pre-trained graph attention network model. During the message aggregation stage, for nodes whose source is the node to be inferred... The message's attention weight is multiplied by a decay factor. : ; in, For the message source node Uncertainty parameter, For the corresponding attenuation factor, This is the ID of the message source node. Uncertainty parameter. Before substituting into the attenuation factor formula, decimal scaling normalization is used to normalize the uncertainty parameter. Mapped to Dimensionless values ​​within the range, ensuring The value of is in It decreases monotonically within the range. When the uncertainty parameter... When it is zero, When the uncertainty parameter is 1, the message weight is unaffected. When it increases, When the value approaches zero, the message contribution of the node to be inferred is suppressed.

[0042] The graph attention network's input layer receives the node feature vectors and adjacency matrices of each node, while the output layer is a fully connected layer that outputs the predicted cascade impact values ​​for each downstream node. During training, the actual impact values ​​of each node in historically known cascade bias events are used as supervision labels. The mean squared error loss function is employed, and the Adam optimization algorithm is used for parameter updates. The training mode is supervised training.

[0043] Step S403: After multiple layers of message passing, the graph attention network outputs the cascading impact prediction values ​​for each downstream node. Based on the cumulative propagation of uncertainty parameters during graph propagation, the prediction confidence interval is calculated. .in, For nodes Cascade impact prediction values, For nodes The boundary value of the prediction uncertainty, This is the downstream node's number, obtained by weighted summation of uncertainty parameters on all upstream message paths of that node.

[0044] Specifically, predicting the boundary value of uncertainty The calculation method is as follows: along the direction of the directed transmission spectrum pointing to the node For all upstream message paths, the sum of the uncertainty parameters of each node on each path is multiplied by the cumulative causal strength of the corresponding path (i.e., the product of the causal strengths of each directed edge on the path). This summation is then divided by the number of paths to obtain the prediction uncertainty boundary value. The uncertainty parameters at each node participate in the prediction uncertainty boundary value. Before calculation, the values ​​are processed into dimensionless values ​​using the same decimal scaling and normalization method as in step S402, and the causal strength is determined. Since the weights themselves are dimensionless, their product and weighted sum are also dimensionless quantities. (This refers to the boundary value of the prediction uncertainty.) The final result, after inverse normalization transformation, is restored to the monetary dimension and compared with the cascading shock prediction value. With consistent dimensions, it can be used in addition and subtraction operations of confidence intervals. The specific formula is: ; in, Pointing to a node The set of all upstream message paths, For one of the paths, For path The directed edge on, For directed edges The causal strength, For path The nodes on For nodes The uncertainty parameter after normalization The total number of paths, This is the number of the downstream node of the target.

[0045] It should be noted that the graph attention network described above introduces an uncertainty decay factor during the message passing phase. The attention weights are adjusted, unlike standard graph attention networks where attention weights are determined solely by node features. An uncertainty decay factor is introduced. Subsequently, node messages with low credibility are automatically suppressed, making the aggregation results more focused on actual observation data.

[0046] In this embodiment, to improve the robustness of cascading impact prediction, the number of message passing layers in the graph attention network is set based on the number of hops in the longest transmission path in the directed transmission graph. When a loop exists in the directed transmission graph, a topological sort is performed on the directed transmission graph and reverse edges are truncated before message passing to ensure that the message passing direction is consistent with the causal transmission direction.

[0047] Step 5, generate budget control signals: The cascading impact forecasts of each node are superimposed on the baseline expenditure forecast. The difference between the superimposed cumulative expenditure forecast and the remaining budget amount is calculated. Based on the relationship between the difference and the forecast confidence interval, a budget control signal of the corresponding level is generated.

[0048] Step 5 may specifically include: Step S501: For each downstream node Obtain its baseline expenditure forecast. The cascading impact predictions of the attention network output. The predicted results under normalized scale are then inversely normalized before being overlaid to be the baseline expenditure predictions. Same monetary units. Cascade shock forecasts. Overlay to baseline expenditure forecast Above, obtain the cumulative expenditure forecast. : ; in, For nodes Expenditure forecasts based on historical trends The predicted values ​​of the cascade impacts are after inverse normalization. This is the cumulative expenditure forecast. These are the numbers for downstream nodes, and all three are measured in monetary units.

[0049] Step S502: Obtain Nodes Remaining budget Calculate the difference between the cumulative expenditure forecast and the remaining budget. : ; in, This is the cumulative expenditure forecast. For the remaining budget, The difference between the two is These are the numbers for downstream nodes, and all three are measured in monetary units.

[0050] Step S503: Difference With budget control threshold The comparisons are made, and the prediction confidence intervals are considered for determination. When And the lower bound of the prediction confidence interval satisfies At that time, a high-confidence budget control signal is generated. Among them, The preset budget control threshold is expressed in monetary units. This refers to the prediction uncertainty boundary value calculated in step S403. The high-confidence budget control signal simultaneously outputs the suggested budget compression ratio and the main transmission link label.

[0051] Step S504: When However, the lower bound of the predicted confidence interval satisfies At that time, a budget warning signal to be confirmed is generated. The budget warning signal to be confirmed is marked with the name of the low-frequency reporting node that causes uncertainty and the expected next data arrival time of the low-frequency reporting node.

[0052] It should be noted that the above-mentioned budget reduction ratio is the percentage by which the cumulative expenditure forecast exceeds the budgeted amount, used to guide subsidiaries to reduce the expenditure plans of the corresponding items in subsequent periods. The above-mentioned main transmission link annotation is obtained by tracing the path from the trigger node to the current node in the directed transmission graph, prioritizing the path with the highest cumulative causal strength as the main transmission link. Main transmission link annotation enables the group headquarters to locate the source subsidiary of the deviation and the downstream subsidiaries most affected.

[0053] In this embodiment of the application, to avoid short-term fluctuations triggering frequent budget control signals, a time window smoothing process can be added to step S503. Specifically, a high-confidence budget control signal is generated only when the high-confidence condition is met in multiple consecutive high-frequency data arrival cycles. If the condition is met only in a single cycle, an observation signal is generated first, and subsequent data confirmation is awaited.

[0054] Step 6: Update the budget control instructions based on actual arrival data: Once the actual data from the low-frequency reporting node arrives, the corresponding posterior estimate is replaced with the actual observation. The local message passing calculation of the graph attention network is re-executed to update the cascaded impact prediction and prediction confidence interval. The generated budget control signal is then corrected or confirmed, and the final budget control execution instruction is output.

[0055] Step 6 may specifically include: Step S601: Receive the actual reported data from the low-frequency reporting node. Extract the actual expenditure deviation value. Replace the posterior estimate of the low-frequency reporting node with the actual observed value. Simultaneously, update the uncertainty parameter of the low-frequency reporting node to zero.

[0056] Step S602: Centered on the data update node, determine the affected local subgraph range in the directed transmission graph. Only re-execute the message passing aggregation operation of the graph attention network on nodes within this local subgraph. Obtain the updated cascading impact prediction values ​​and prediction confidence intervals.

[0057] Step S603: Compare the updated cascaded impact prediction value with the generated budget control signal. If the updated prediction result still meets the high-confidence control conditions, confirm the original budget control signal and output the final budget control execution command. If the updated prediction result no longer meets the control conditions, cancel the original budget control signal and output a correction notification. If the original budget control signal is a budget warning signal to be confirmed and the updated signal meets the high-confidence conditions, upgrade the budget warning signal to a high-confidence budget control signal and output the corresponding budget control execution command.

[0058] In this embodiment, to reduce the computational overhead of the local update process, the scope of the local subgraph in step S602 is determined as follows: starting from the data update node, traversing along the outgoing edges of the directed transmission graph, only downstream nodes whose shortest path hop count to the data update node does not exceed a preset hop count limit are included. This processing method maintains update accuracy while avoiding re-performing the complete message passing calculation on the entire directed transmission graph.

[0059] On the other hand, the present invention also proposes a financial budget control system based on big data analysis, comprising: The data acquisition and graph generation module is used to acquire expenditure execution data streams and internal transaction flow data streams of budget items of each subsidiary of the group, register update frequency tags and latest arrival timestamps for each expenditure execution data stream, perform Granger causality tests based on internal transaction flow data, and generate directed transmission graphs between budget items of the group.

[0060] The cascading trigger detection module is used to calculate the ratio of expenditure deviation change rate to historical volatility when a new data event is generated at a high-frequency reporting node. When the ratio exceeds a preset multiple threshold, it is marked as a cascading trigger event and the correlation analysis process is activated.

[0061] The low-frequency node state inference module is used to calculate the posterior estimate and uncertainty parameter of the expenditure deviation of the low-frequency reporting downstream node at the current time for the low-frequency reporting downstream node in the directed transmission spectrum that has not yet received new data, based on the conditional probability distribution of the transmission relationship, the transmission delay parameter and the deviation sequence observed by the upstream node.

[0062] The cascade impact prediction module is used to take the actual observations or posterior estimates of each node as node feature vectors after Z-score standardization, and use the uncertainty parameters after decimal scaling normalization as attention weight adjustment factors. The input graph attention network performs message passing aggregation operations along the directed transmission graph, and outputs the cascade impact prediction values ​​and prediction confidence intervals of each downstream node.

[0063] The budget control signal generation module is used to superimpose the cascading impact forecast value after inverse normalization processing onto the baseline expenditure forecast value, calculate the difference between the cumulative expenditure forecast value and the remaining budget amount, and generate a high-confidence budget control signal or a budget warning signal to be confirmed based on the relationship between the difference and the forecast confidence interval.

[0064] The data update and instruction output module is used to replace the corresponding posterior estimate after the actual data arrives at the low-frequency reporting node, re-execute the local message passing calculation, update the cascade impact prediction value and prediction confidence interval, correct or confirm the generated budget control signal, and output the final budget control execution instruction.

[0065] By way of example, the system proposed in this invention may also include various features and combinations thereof in the method embodiments, which will not be elaborated here.

[0066] The beneficial effects of this invention are as follows: This invention extracts directed transmission graphs between budget items from internal transaction flow data using Granger causality tests, enabling big data analytics to trace the transmission path of cascading deviations along the directional causal relationships of the internal supply chain. Therefore, it overcomes the limitations of traditional independent analysis models where subsidiary budget data is processed in isolation and cross-subsidiary cascading effects cannot be perceived.

[0067] By using conditional probability posterior estimation to perform real-time state inference for downstream nodes in low-frequency reporting, the graph attention network can perform cascade propagation analysis without waiting for all node data to align. Therefore, the response time of the budget control signal is advanced from waiting for the complete arrival of low-frequency reporting node data to the occurrence of high-frequency data events, eliminating the weeks-long analysis delay caused by differences in reporting frequencies.

[0068] By introducing an uncertainty decay factor into message aggregation in a graph attention network, the message contribution of posterior estimated nodes is automatically adjusted by the inference confidence level. Therefore, it avoids overestimating or underestimating the strength of the transmission effect when directly using historical data to fill in the gaps, ensuring that the accuracy of budget control decisions is not significantly reduced due to missing data from some nodes.

[0069] Based on the above improvements, budget control decisions are guaranteed in both timeliness and accuracy, enabling precise financial budget control based on big data correlation analysis in scenarios of cascading deviations within the internal supply chain of group enterprises.

[0070] See Figures 2-9 The following application examples are proposed: A large manufacturing conglomerate consists of three tiers of supply chain subsidiaries. Upstream Company A is responsible for raw material procurement, midstream Company B is responsible for manufacturing, and downstream Company C is responsible for product sales. Company A reports budget execution data to the group headquarters daily, Company B weekly, and Company C monthly. In quarter A of a certain year, global commodity prices fluctuated significantly, causing a persistent expenditure deviation in Company A's procurement account. The group headquarters needs to assess the cascading risk of this deviation affecting Company B's manufacturing account and Company C's sales account, in order to issue budget control signals in advance before downstream data arrives.

[0071] The group's big data analytics platform collects expenditure execution data streams from three subsidiaries, as well as internal transaction records, extracting raw material settlement records from Company A to Company B and semi-finished product transfer records from Company B to Company C. Based on these transaction amount time series, the platform performs Granger causality tests pairwise on the expenditure time series of the three accounts. The test results confirm that Company A's purchasing account has a significant predictive ability for Company B's manufacturing account, with the optimal lag order corresponding to a two-week transmission delay; Company B's manufacturing account has a significant predictive ability for Company C's sales account, with the optimal lag order corresponding to a three-week transmission delay. The resulting directed transmission graph contains two directed edges, with the directions being Company A's purchasing account pointing to Company B's manufacturing account, and Company B's manufacturing account pointing to Company C's sales account, respectively.

[0072] Table 1. Edge properties of directed transmission graphs

[0073] On Monday of the eighth week of quarter A in a certain year, Company A pushes out the daily expenditure execution data for its procurement account. The platform calculates the difference between the actual expenditure and the budgeted value for the day to obtain the current expenditure deviation value; then it calculates the difference between the current expenditure deviation value and the previous day's expenditure deviation value to obtain the expenditure deviation change rate. Simultaneously, extract the historical deviation sequence of Company A's procurement items for the past thirty trading days and calculate the historical volatility. According to the formula Calculate the deviation trigger ratio , and the preset multiple threshold After comparison, the determination is made. This triggers a cascading event, activating the correlation analysis process.

[0074] Table 2. Triggering Test Data for Company A's Procurement Items

[0075] After the association analysis process is activated, the platform traverses the downstream nodes of Company A's purchasing category along the outgoing edges of the directed transmission graph. Company B's manufacturing category has its latest arrival timestamp as last Friday, which is more than one business day from the current analysis time. No new data has been submitted in the current analysis window, so it is marked as a node to be inferred. Company C's sales category has its latest data submitted at the end of last month, and is also marked as a node to be inferred.

[0076] For Company B's manufacturing account, its only updated upstream parent node is Company A's purchasing account. Extract the observed deviation value of Company A's purchasing account two weeks prior to the propagation delay. Combining causal strength Calculate the a posteriori estimate of Company B's manufacturing account using the weighted summation formula. Since there is only one upstream parent node, the posterior estimate, after weighted summation and dividing by the sum of weights, is equal to Company A's observed deviation at that moment. Uncertainty parameter. The variance is calculated based on the historical linear regression residuals. The upstream parent node of Company C's sales account is Company B's manufacturing account, and Company B's manufacturing account itself is also a node to be inferred. Therefore, the posterior estimate of Company B's manufacturing account is used in the calculation, and the uncertainty parameter of Company B's manufacturing account is added to the uncertainty parameter of Company C.

[0077] Table 3. Posterior inference results for low-frequency nodes

[0078] The platform performs Z-score standardization on the expenditure deviations of the three nodes to eliminate differences in expenditure scale among subsidiaries. For Company A's procurement item, the actual observed deviation value is used, with the uncertainty parameter set to zero; for Companies B and C, the posterior estimates are used, with corresponding uncertainty parameters added. The standardized node feature vectors, along with their adjacency matrices, are input into a pre-trained graph attention network. During the message aggregation phase, the attenuation factor for Company B's manufacturing item is... according to The calculation shows that the normalized uncertainty parameter suppresses the attention weight of Company B's node and further suppresses the attention weight of Company C's node. After two layers of message passing, the attention network outputs the cascading impact prediction values ​​of each downstream node and calculates the prediction confidence interval.

[0079] Table 4. Output results of the attention network.

[0080] The platform generates budget control signals for Company B's manufacturing account and Company C's sales account respectively. The cascading impact forecasts, after inverse normalization, are superimposed onto the baseline expenditure forecasts for each node to obtain the cumulative expenditure forecast. Then, with the remaining budget The difference is obtained by subtracting. With budget control threshold The difference in Company B's manufacturing account is compared to the previous one, and it is verified whether the lower bound of the predicted confidence interval exceeds the remaining budget. If the threshold is exceeded and the lower bound of the confidence interval still exceeds the remaining budget, the high-confidence control condition is met. A high-confidence budget control signal is generated, and a suggested budget reduction ratio and the main transmission link label (Company A's purchasing account → Company B's manufacturing account) are output. The difference between Company C's sales account and the remaining budget is also output. If the threshold is exceeded, but the lower bound of the confidence interval does not exceed the remaining budget, a budget warning signal pending confirmation is generated. The low-frequency node causing uncertainty is marked as the manufacturing item of Company B. The next data arrival time is expected to be this Friday.

[0081] Table 5. Results of Budget Control Signal Generation

[0082] This Friday, the actual data reported by Company B's manufacturing account arrived. The platform extracted the actual expenditure deviation value of Company B's manufacturing account, replaced the posterior estimate with the actual observed value, and updated the uncertainty parameter to zero. Centering on Company B's manufacturing account, the platform determined the affected local subgraph range and re-executed local message passing aggregation operations only on Company B's manufacturing account and its downstream Company C's sales account. After the update, the cascading impact prediction value of Company B's manufacturing account was recalculated based on the actual observed value, and the prediction confidence interval narrowed. The uncertainty parameter of Company C's sales account decreased due to the deterministic update of Company B's node, and the lower bound of the confidence interval shifted upward, meeting the high-confidence control conditions after assessment. The platform upgraded the pending budget warning signal for Company C's sales account to a high-confidence budget control signal and output the corresponding budget control execution instruction.

[0083] Table 6. Signal Correction Results After Data Update

[0084] The entire data flow process presents a complete causal chain. Step 1 extracts the directed transmission graph from Company A's purchasing account to Company B's manufacturing account and then to Company C's sales account from the internal transaction flow, laying the structural foundation for subsequent cascade analysis. Step 2 triggers cascade analysis as soon as Company A's high-frequency data arrives, without waiting for Company B and Company C's low-frequency data. Step 3 uses Company A's observed deviation value to infer the current deviation state of Company B and Company C along the transmission delay, and generates uncertainty parameters for the two low-frequency nodes. Step 4 inputs the deviation values ​​and uncertainty parameters of the three nodes into a graph attention network. The uncertainty attenuation factor differentially suppresses the message contribution of the low-frequency nodes, outputting cascade impact prediction values ​​with confidence intervals. Step 5 generates a high-confidence control signal for Company B and a pending confirmation warning signal for Company C based on the combination relationship between the difference and the confidence interval. Step 6 performs a partial update after Company B's actual data arrives, upgrading Company C's pending confirmation warning signal to a high-confidence control signal. The entire data link completes the early warning loop from the moment Company A triggers the high-frequency data. Compared with the traditional batch analysis model that requires waiting for Company C's monthly data, this effectively reduces the response delay of cascading deviations.

[0085] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A financial budget control method based on big data analysis, characterized in that, include: Obtain expenditure execution data streams and internal transaction flow data streams for each subsidiary's budget items. Perform Granger causality tests on the expenditure time sequence of each subsidiary's budget items based on the internal transaction flow data to generate a directed transmission graph between budget items. When the ratio of the expenditure deviation change rate of the high-frequency reporting node to the historical volatility exceeds a preset multiple threshold, it is marked as a cascading trigger event. For low-frequency downstream nodes in the directed transmission spectrum that have not yet received new data, the posterior estimate and uncertainty parameter are calculated based on the conditional probability distribution of the transmission relationship and the observed deviation sequence of the upstream node. The actual observed deviation or posterior estimated value of each node is used as the node feature vector input to the graph attention network. During the message aggregation stage, the attention weight of the posterior estimated node is attenuated and adjusted according to the uncertainty parameter, and the cascade impact prediction value and prediction confidence interval of each downstream node are output. The cascading shock forecasts are superimposed on the baseline expenditure forecasts, and a budget control signal is generated based on the difference between the cumulative expenditure forecasts and the remaining budget amount, as well as the forecast confidence interval.

2. The financial budget control method based on big data analysis according to claim 1, characterized in that, In the directed transmission graph, nodes represent the budget items of each subsidiary, directed edges represent the transmission direction, edge weights are the statistic values ​​of the Granger causality test as the causal strength, and edge attributes include a transmission delay parameter, which is the time interval corresponding to the optimal lag order in the Granger causality test.

3. The financial budget control method based on big data analysis according to claim 1, characterized in that, The events marked as cascading triggers include: The difference between the actual expenditure value and the budgeted expenditure value for the current period at the high-frequency reporting node is used to obtain the current expenditure deviation value. The difference between the current expenditure deviation value and the expenditure deviation value of the previous period is used to obtain the expenditure deviation change rate. ; Obtain the standard deviation of the expenditure deviation sequence within the historical period of this node as the historical volatility. ; Calculate the deviation trigger ratio ,when The time marker is designated as a cascading event; where, For nodes The rate of change in expenditure deviation, For nodes Historical volatility, For the preset multiple threshold, Number the nodes.

4. The financial budget control method based on big data analysis according to claim 1, characterized in that, The calculation of the posterior estimate and uncertainty parameters includes: Traverse the directed transmission graph from the node that triggered the cascading event, marking nodes that have not reported new data within the current analysis window as nodes to be inferred; for each node to be inferred... Get the set of all its updated upstream parent nodes. Extract each upstream parent node In conduction delay Previous observed deviation values According to the strength of causation Weighted posterior estimate: ;in, For the current analysis time, For nodes To node The propagation delay parameter, For nodes To node The causal strength; Uncertainty parameter Based on the deviations of each upstream parent node in historical data and the node to be inferred The residual variance is obtained by calculating the conditional distribution between deviations.

5. The financial budget control method based on big data analysis according to claim 4, characterized in that, When the upstream parent node of the node to be inferred is also the node to be inferred, the posterior estimate of the upstream parent node is used to replace the actual observation value in the calculation, and the uncertainty parameter of the upstream parent node is added to the uncertainty parameter of the current node to be inferred; when the time interval between the latest data of the low-frequency reported downstream node and the current time exceeds the preset multiple of the reporting period of the node, the uncertainty parameter of the node is set to the upper limit value.

6. The financial budget control method based on big data analysis according to claim 1, characterized in that, The attenuation adjustment of the attention weights for the posterior estimation nodes according to the uncertainty parameter includes: For nodes whose source is to be inferred The message's attention weight is multiplied by a decay factor. ,in For the message source node Uncertainty parameters after normalization; when When the attenuation factor is zero, it is 1. As the factor increases, the attenuation factor approaches zero.

7. The financial budget control method based on big data analysis according to claim 6, characterized in that, The prediction confidence interval is: Among them, the boundary value of prediction uncertainty The calculation method is as follows: ,in Pointing to a node The set of all upstream message paths, For one of the paths, For directed edges on the path The causal strength, For nodes on the path The uncertainty parameter after normalization The total number of paths, For nodes The predicted value of cascading impacts.

8. The financial budget control method based on big data analysis according to claim 1, characterized in that, The generated budget control signal includes: Calculate the cumulative expenditure forecast With remaining budget The difference ;when Exceeding the budget control threshold And the lower bound of the prediction confidence interval satisfies At that time, a high-confidence budget control signal is generated; when Exceeding the budget control threshold However, the lower bound of the prediction confidence interval is satisfied. At that time, a budget warning signal awaiting confirmation is generated; among which For nodes Baseline expenditure forecasts, The predicted values ​​of the cascade impacts are after inverse normalization. To predict the uncertainty boundary values, Number the downstream node.

9. The financial budget control method based on big data analysis according to claim 1, characterized in that, Also includes: When the actual data from the low-frequency reporting node arrives, the posterior estimate of that node is replaced with the actual observation and the uncertainty parameter is updated to zero. The affected local subgraph range is determined with the data update node as the center. The message passing aggregation operation of the graph attention network is re-executed only for the nodes within the local subgraph. The cascaded impact prediction value and prediction confidence interval are updated. The generated budget control signal is corrected or confirmed, and the final budget control execution instruction is output.

10. A financial budget control system based on big data analytics, used to execute the financial budget control method based on big data analytics as described in any one of claims 1 to 9, characterized in that, include: The data acquisition and graph generation module is used to acquire expenditure execution data flow and internal transaction flow data flow of budget items of various subsidiaries of the group, perform Granger causality test based on internal transaction flow data, and generate directed transmission graph between budget items; The cascading trigger detection module is used to calculate the ratio of the expenditure deviation change rate to the historical volatility when a new data event is generated at the high-frequency reporting node. When the ratio exceeds a preset multiple threshold, it is marked as a cascading trigger event. The low-frequency node state inference module is used to calculate the posterior estimate and uncertainty parameters for low-frequency downstream nodes in the directed transmission spectrum that have not yet received new data, based on the conditional probability distribution of the transmission relationship and the observed deviation sequence of the upstream node. The cascaded impact prediction module is used to input the actual observed deviation value or posterior estimated value of each node as the node feature vector into the graph attention network. During the message aggregation stage, the attention weight of the posterior estimated node is attenuated and adjusted according to the uncertainty parameter, and the cascaded impact prediction value and prediction confidence interval of each downstream node are output. The budget control signal generation module is used to superimpose the cascading shock forecast value onto the baseline expenditure forecast value, and generate a budget control signal based on the difference between the cumulative expenditure forecast value and the remaining budget amount and the forecast confidence interval.

Images