Budget performance and fund use linkage analysis and closed-loop management method

Abstract

The present application relates to the field of financial management, and discloses a budget performance and fund use linkage analysis and closed-loop management method, comprising the following steps: obtaining multi-source heterogeneous time series data in the budget project execution process to form a standardized multi-dimensional time series feature matrix; and based on the dynamic causal diagram, estimating a structural equation model reflecting the causal relationship between fund variables and performance variables; and combining the intervention weight to generate a root cause ranking list; constructing a mixed extreme value distribution model and quantifying the additional risk increment caused by each root cause; constructing a multi-objective optimization model to generate a resource redistribution scheme; and tracking the execution effect of the control instruction. Through the fusion of causal inference and counterfactual attribution, the root cause affecting performance is accurately located, the extreme risk of management anchoring is quantified based on the extreme value theory, and the resource redistribution decision is generated based on the multi-objective optimization, so that a complete management closed loop from intelligent analysis, risk warning to optimization control and feedback calibration is constructed.

Description

Technical Field

[0001] This invention relates to the field of financial management technology, specifically to a method for linking budget performance with fund utilization and for closed-loop management. Background Technology

[0002] With the popularization of information technology, various financial accounting systems, business management systems, and data visualization tools are widely used to improve the precision and scientific level of management. Current common practice is to treat the recording of budgeted fund usage and the evaluation of project performance as two relatively independent management threads. On the one hand, the financial system is responsible for recording the flow of funds allocation, expenditure, and settlement; on the other hand, the performance management system scores and evaluates projects based on a pre-set performance indicator system after the project cycle ends or at fixed points.

[0003] Existing analyses linking budget performance and fund usage often rely on simple judgments based on static thresholds of single indicators or visualizations of multi-dimensional data relationships. This makes it impossible to model and diagnose the inherent causal mechanisms between cash flow and performance flow. Furthermore, risk quantification is often based on assumptions about the overall data distribution, making it difficult to accurately estimate the probability and assess the losses of extreme risk events that management is concerned about. Consequently, the intelligence and accuracy of analysis and early warning are insufficient. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method for linking budget performance with fund utilization analysis and closed-loop management, which solves the problems of insufficient intelligence and accuracy in existing linking analysis and early warning systems for budget performance and fund utilization.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for linking budget performance with fund utilization and for closed-loop management, comprising the following steps: Acquire multi-source heterogeneous time-series data during the execution of budget projects, and fuse and perform feature engineering on the multi-source heterogeneous time-series data to form a standardized multi-dimensional time-series feature matrix; Based on the multidimensional time-series feature matrix, a dynamic causal graph is constructed by integrating prior knowledge of the domain, and a structural equation model reflecting the causal relationship between funding variables and performance variables is estimated based on the dynamic causal graph. Based on the structural equation model, counterfactual reasoning and attribution analysis are performed on the detected performance anomalies to obtain performance deviation data, calculate the average treatment effect of each candidate cause, and generate a root cause ranking list by combining interventionist weights. Based on historical and current performance deviation data, a hybrid extreme value distribution model is constructed to calculate the tail risk probability and conditional risk value under a preset management anchor threshold, and to quantify the additional risk increment caused by each fundamental cause. Based on the fundamental cause ranking list, the tail risk probability and conditional risk value, and the additional risk increment, a multi-objective optimization model is constructed to generate a resource reallocation scheme, and the resource reallocation scheme is transformed into an executable control instruction. The execution effect of the control instructions is tracked, and the parameters of the dynamic causal graph, the structural equation model, and the mixed extreme value distribution model are calibrated based on the execution effect.

[0006] By adopting the above technical solutions, the fundamental factors affecting performance are accurately located by integrating causal inference and counterfactual attribution, and extreme risks anchored by management are quantified by combining extreme value theory. Based on multi-objective optimization, resource reallocation decisions are generated, thus constructing a complete management closed loop from intelligent analysis and risk warning to optimization control and feedback calibration. This achieves deep perception, accurate diagnosis and adaptive control of the budget execution process, and solves the problems of insufficient intelligence and accuracy in the existing linkage analysis of budget performance and fund use.

[0007] Preferably, the formation of the standardized multidimensional time-series feature matrix includes the following steps: Real-time acquisition of cash flow data, performance indicator data, and environmental context data, and alignment and interpolation at a uniform time granularity; Calculate budget execution rate, expenditure volatility and payment concentration characteristics from cash flow data; calculate achievement speed and quality stability characteristics from performance indicator data; and extract policy impact factors and market volatility index characteristics from environmental context data through natural language processing and numerical calculation. All extracted features are organized according to time series to form a multidimensional time series feature matrix with uniform dimensions.

[0008] Preferably, the step of constructing a dynamic causal graph based on the multidimensional temporal feature matrix and fusing domain prior knowledge includes the following steps: The prior knowledge in the field of budget management is formalized into time-irreversible constraints, rigid budget transmission constraints, and lagging institutional and process constraints, thus forming prior knowledge constraints. Within the sliding time window, based on the prior knowledge constraints, an improved causal inference algorithm is executed on the multidimensional time series feature matrix to perform conditional independence testing and causal direction determination. The output part is a directed acyclic graph as the causal structure of the current window, and the stability of the structure is evaluated by calculating the edit distance of the causal graphs of adjacent windows. The structure is updated when it changes, forming a dynamic causal graph.

[0009] Preferably, the step of estimating a structural equation model reflecting the causal relationship between funding variables and performance variables based on the dynamic causal graph includes the following steps: For each variable in the dynamic causal graph, the structural equation coefficients are estimated using a linear regression method with signed constraints, with its parent node as the independent variable. The structural equation coefficients are used to construct linear equations for each variable, forming a structural equation model that characterizes the quantitative causal relationship between the variables.

[0010] Preferably, the step of performing counterfactual reasoning and attribution analysis on the detected performance anomalies to obtain performance deviation data and calculate the average treatment effect of each candidate cause includes the following steps: Locate all candidate cause nodes leading to abnormal performance variables in the dynamic cause-effect graph; For each candidate cause, within the framework of the structural equation model, an intervention is performed to set the cause variable to its planned value or normal state value; Calculate the counterfactual performance expectation after intervention, and use the difference between the observed performance and the counterfactual performance expectation as the average treatment effect of the candidate cause; The difference between the observed performance and the expected counterfactual performance is recorded as the performance deviation data for the current monitored event.

[0011] Preferably, the step of generating the root cause ranking list by combining the interventionibility weights includes: Assign a weight coefficient representing the manageability of each candidate cause; The absolute value of the average treatment effect of each candidate cause after time-adjusted is normalized to obtain the original contribution. The original contribution score is multiplied by the corresponding interventionist weight coefficient and then normalized twice to obtain the ranking score. Candidate causes are sorted in descending order of their ranking scores to generate a root cause ranking list.

[0012] Preferably, the construction of the hybrid extreme value distribution model to calculate the tail risk probability and conditional value of risk under a preset management anchoring threshold includes the following steps: Performance deviation is defined as the relative deviation between the actual value and the target value. Suppose that the performance deviation follows a mixed distribution consisting of a normal distribution and a generalized Pareto distribution; The weights, normal distribution parameters, and generalized Pareto distribution parameters of the mixed distribution are estimated using the expectation-maximization algorithm to form a mixed distribution model; Based on the mixed distribution model, the cumulative probability of performance deviation being lower than each preset management anchor threshold is calculated as the tail risk probability; The expected value of performance deviation is calculated as the conditional value of risk, given that the performance deviation is already below the management anchor threshold.

[0013] Preferably, the quantification of the additional risk increment caused by each fundamental cause includes: Based on the reasons in the root cause ranking list, the historical data is divided into two groups: one group where the reason is in an abnormal state and the other group where it is in a normal state. The tail risk probability is calculated based on the two sets of data respectively; The difference between the two sets of tail risk probabilities is taken as the additional risk increment caused by the fundamental cause.

[0014] Preferably, the step of constructing a multi-objective optimization model to generate a resource reallocation scheme includes the following steps: The resource allocation for each budget item is used as the decision variable; The optimization objective is to maximize the total expected performance improvement of all projects and minimize the total extreme risks of all projects. The constraints are the total adjustable amount of each budget item and the basic resource requirements for project operation; Based on the optimization objectives and constraints, a multi-objective optimization function is constructed and solved to obtain the Pareto optimal solution as a resource reallocation scheme, and the expected performance improvement value corresponding to the scheme is recorded.

[0015] Preferably, the step of tracking the execution effect of the control instructions and performing feedback calibration on the parameters of the dynamic causal graph, the structural equation model, and the mixed extreme value distribution model based on the execution effect includes the following steps: Calculate the actual performance improvement within a preset time window after the execution of control instructions; Compare the actual degree of performance improvement with the expected performance improvement value; When the actual improvement reaches the expected performance improvement value, the causal coefficient of the path related to this intervention in the dynamic causal diagram will be increased proportionally. If the actual improvement does not reach the expected improvement value, the relevant causal structure is rediscovered and reassessed. The current control order, the corresponding resource reallocation plan, and the actual degree of performance improvement are used as new samples and added to the historical database to retrain the hybrid extreme value distribution model.

[0016] This invention provides a method for linking budget performance with fund utilization analysis and closed-loop management. It has the following beneficial effects: 1. This invention accurately identifies the root causes affecting performance by integrating causal inference and counterfactual attribution, quantifies extreme risks anchored by management using extreme value theory, and generates resource reallocation decisions based on multi-objective optimization. This constructs a complete management closed loop from intelligent analysis and risk warning to optimization control and feedback calibration, achieving deep perception, accurate diagnosis, and adaptive regulation of the budget execution process. It solves the problems of insufficient intelligence and accuracy in existing budget performance and fund usage linkage analysis and early warning.

[0017] 2. This invention introduces dynamic causal graphs and structural equation modeling to formally model and continuously update the causal relationships between cash flow, performance flow, and environmental factors in the budget execution process. It uses counterfactual reasoning techniques to perform quantitative attribution analysis on monitored performance anomalies, which can accurately locate the root causes of the anomalies and their specific contributions, and generate a priority ranking list that considers management intervention, thereby improving the intelligence level of early warning analysis and the interpretability of diagnostic results.

[0018] 3. This invention models performance deviation data using a hybrid extreme value distribution model, separating regular fluctuations from extreme risks. By calculating the tail risk probability and conditional risk value under a preset management anchor threshold, it provides management with graded and quantifiable risk measurement indicators. By quantifying the additional contribution of each fundamental cause to tail risk, it establishes a quantitative correlation between risk events and specific management issues. This enables risk warnings to not only indicate the magnitude of the risk but also reveal its source, thereby achieving a shift from vague perception to precise quantification and improving the pertinence and foresight of risk management. Attached Figure Description

[0019] Figure 1 The flowchart shows the budget performance and fund utilization linkage analysis and closed-loop management method proposed in this invention. Figure 2 This is an architecture diagram of the budget performance and fund utilization linkage analysis and closed-loop management system proposed in an embodiment of the present invention. Detailed Implementation

[0020] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Example 1: In the first embodiment of the present invention, the present invention provides a method for linking budget performance with fund utilization analysis and closed-loop management, such as... Figure 1 As shown, it includes the following steps: Acquire multi-source heterogeneous time-series data during the execution of budget projects, and fuse and perform feature engineering on the multi-source heterogeneous time-series data to form a standardized multi-dimensional time-series feature matrix; Furthermore, a standardized multidimensional time-series feature matrix is ​​formed, including the following steps: Real-time acquisition of cash flow data, performance indicator data, and environmental context data, and alignment and interpolation at a uniform time granularity; Calculate budget execution rate, expenditure volatility and payment concentration characteristics from cash flow data; calculate achievement speed and quality stability characteristics from performance indicator data; and extract policy impact factors and market volatility index characteristics from environmental context data through natural language processing and numerical calculation. All extracted features are organized according to time series to form a multidimensional time series feature matrix with uniform dimensions.

[0022] Specifically, a standardized multi-dimensional time-series feature matrix is ​​formed through multi-source data synchronization, time alignment, and feature construction. Three types of heterogeneous data—fund flow, performance indicators, and environmental context—are acquired in real time, and time-series alignment and missing value interpolation are performed at a unified daily granularity to ensure consistent data time reference.

[0023] Specifically, for aligned funding data, features such as budget execution rate, expenditure volatility, and payment concentration are extracted. In some embodiments, payment concentration is calculated using the Herfindahl-Hirschman Index (HHI). ,in Indicates time Payment to the recipient The amount, This represents the total payment amount at that moment. This formula quantifies the concentration of fund flows, and the result serves as the input to the subsequent matrix as a feature of payment concentration.

[0024] The system extracts speed and quality stability characteristics from performance data; for environmental data, it extracts policy influencing factors through natural language processing and calculates market volatility indices through numerical computation. All features are organized chronologically into a standardized matrix with rows corresponding to time points and columns corresponding to feature dimensions. This matrix provides a structured and unified data foundation for subsequent causal discovery and modeling, supporting cross-modal deep correlation analysis.

[0025] Based on a multidimensional time-series feature matrix, a dynamic causal graph is constructed by integrating prior knowledge of the domain, and a structural equation model reflecting the causal relationship between funding variables and performance variables is estimated based on the dynamic causal graph. Furthermore, based on a multi-dimensional temporal feature matrix, a dynamic causal graph is constructed by fusing domain prior knowledge, including the following steps: The prior knowledge in the field of budget management is formalized into time-irreversible constraints, rigid budget transmission constraints, and lagging institutional and process constraints, thus forming prior knowledge constraints. Within the sliding time window, under the premise of prior knowledge constraints, an improved causal inference algorithm is executed on the multi-dimensional time series feature matrix to perform conditional independence testing and causal direction determination. The output part is a directed acyclic graph as the causal structure of the current window, and the stability of the structure is evaluated by calculating the edit distance of the causal graphs of adjacent windows. The structure is updated when it changes, forming a dynamic causal graph.

[0026] Furthermore, based on dynamic causal graph estimation, a structural equation model reflecting the causal relationship between funding variables and performance variables is constructed, including the following steps: For each variable in the dynamic causal graph, the structural equation coefficients are estimated using a linear regression method with signed constraints, with its parent node as the independent variable. By using structural equation coefficients, linear equations for each variable are constructed, forming a structural equation model that characterizes the quantitative causal relationship between variables.

[0027] Specifically, the construction of dynamic causal graphs and the estimation of structural equation models enable the quantitative analysis of causal relationships between funding and performance variables. First, prior knowledge in the field of budget management is formalized into constraints that can be processed by the algorithm. The time irreversibility constraint requires that funding flow events cannot be the result of subsequent performance output events; the rigid transmission constraint of the budget limits the impact of the cumulative expenditure rate on marginal performance output to a negative direction; and the institutional process lag constraint defines the minimum time interval between funding disbursement and performance response. These constraints collectively constitute the set of prior knowledge constraints.

[0028] Within a sliding time window, an improved causal inference algorithm is performed on the multidimensional time-series feature matrix, based on this constraint set. One possible approach is to employ a constraint-based fast causal inference algorithm. The core of this process is a conditional independence test to determine whether two features are independent given certain other features. For example, for any two funding or performance variables in the feature matrix... and and a set of condition variables This can be tested by calculating the partial correlation coefficient: ,in, For variables and The correlation coefficient, For variables With condition set The vector of correlation coefficients among all variables in the vector. For condition set The inverse of the correlation coefficient matrix of the internal variables. This formula takes the correlation coefficient matrix between variables as input and outputs the result in the control condition set. Partial correlation coefficient after influence If its absolute value is lower than the statistical significance threshold, then it is determined that... and In a given The condition is independent. Based on this, and combined with the aforementioned domain constraints, the causal direction is determined, and finally a partially directed acyclic graph is output as the causal structure identified within the current time window.

[0029] To address the potential evolution of business logic over time, structural stability is assessed by calculating the graph edit distance between causal graphs obtained from adjacent time windows. When the distance exceeds a preset threshold, a structural change is identified, and the causal graph is updated accordingly, thereby enabling the dynamic evolution of the causal model.

[0030] Based on the finalized dynamic causal graph, the structural equation model can be estimated. For any variable in the causal graph... Using all its parent node variables as independent variables, a linear regression method with sign constraints (e.g., the regression coefficients must conform to the rigid transmission constraint of the budget) is applied to fit the equation, resulting in a set of structural equation coefficients. Using these coefficients, a linear equation of the form: ,in, For variables In the set of all parent nodes in the graph, the coefficient Quantized the parent node child nodes The direct impact on strength, This represents the error term. These equations for all variables collectively constitute a structural equation model, thus quantitatively characterizing the causal mechanism between financial and performance variables, providing a computable mathematical model basis for subsequent counterfactual reasoning.

[0031] Based on structural equation modeling, counterfactual reasoning and attribution analysis are performed on the detected performance anomalies to obtain performance deviation data, calculate the average treatment effect of each candidate cause, and generate a root cause ranking list by combining interventionist weights. Furthermore, counterfactual reasoning and attribution analysis are performed on the detected performance anomalies to obtain performance deviation data, and the average treatment effect of each candidate cause is calculated, including the following steps: Locate all candidate cause nodes leading to abnormal performance variables in the dynamic cause-effect graph; For each candidate cause, within the framework of structural equation modeling, an intervention is performed to set the cause variable to its planned value or normal state value; Calculate the counterfactual performance expectation after intervention, and use the difference between the observed performance and the counterfactual performance expectation as the average treatment effect of the candidate cause; The difference between observed performance and counterfactual performance expectations is recorded as performance deviation data for the current monitored event.

[0032] Furthermore, a ranking list of root causes is generated by combining interventionist weights, including: Assign a weight coefficient representing the manageability of each candidate cause; The absolute value of the average treatment effect of each candidate cause after time-adjusted is normalized to obtain the original contribution. The original contribution score is multiplied by the corresponding interventionist weight coefficient and then normalized twice to obtain the ranking score. Sort candidate reasons in descending order of their scores to generate a sorted list of root causes.

[0033] Specifically, counterfactual reasoning and attribution analysis are used to identify the root causes of performance anomalies and quantify their impact when such anomalies are detected. This analysis process is initiated when the actual observed value of a performance variable Y deviates significantly from the expected target.

[0034] Specifically, firstly, a reverse causal analysis is performed based on the established dynamic causal graph to locate all nodes in the causal path that directly or indirectly affect the abnormal performance variable. The various characteristic variables corresponding to these nodes are then identified as a set of candidate causes. For each candidate cause variable in the set, a counterfactual intervention calculation is performed within the structural equation model framework. The value of the cause variable in the model is modified from its current abnormal observation value to its value under normal planning conditions, while keeping the values ​​or relationships of all other variables in the model unchanged. Subsequently, based on this modified model, the expected value of the performance variable Y is recalculated to obtain the counterfactual performance expectation. .

[0035] Next, the observation performance is calculated. Counterfactual performance expectations The difference between the two is the average treatment effect of the candidate cause. The calculation formula is as follows: ,in, These are the actual observed values ​​of performance variables in performance anomaly events. These are the estimated values ​​of performance variables calculated using structural equation modeling after counterfactual intervention targeting the causes. The numerical values ​​quantify the expected performance loss that could be recovered if the cause were corrected to a normal state. Simultaneously, the counterfactual discrepancies calculated for the primary candidate causes are recorded as performance deviation data characterizing the severity of this anomaly.

[0036] After obtaining the average treatment effect of all candidate causes, a comprehensive ranking is further conducted based on management practices. Each candidate cause is assigned a pre-defined intervention weight, which is set according to the degree of management control over the causal variable. For example, internal factors that can be directly adjusted by the current management level are given a higher weight; factors requiring approval from higher levels or constrained by the external environment are given a lower weight.

[0037] Next, the absolute value of the average treatment effect of all candidate causes is normalized to obtain the raw contribution. Then, the raw contribution is multiplied by the corresponding interventionibility weight to obtain the weighted contribution. Finally, the weighted contribution is normalized a second time to calculate the final ranking score for each cause. In some embodiments, the sorting score The calculation can be expressed as: This calculation process comprehensively weighs the impact of each cause on the abnormal results. And its operability in actual management. Ultimately, it is based on the ranking score. Arrange the results in descending order to generate a root cause ranking list, thus providing managers with a clear and actionable reference for action priorities, directing limited management resources to the most effective and feasible steps.

[0038] Based on historical and current performance deviation data, a hybrid extreme value distribution model is constructed to calculate the tail risk probability and conditional risk value under a preset management anchor threshold, and to quantify the additional risk increment caused by each fundamental cause. Furthermore, a mixed extreme value distribution model is constructed to calculate the tail risk probability and conditional value of risk under a preset management anchor threshold, including the following steps: Performance deviation is defined as the relative deviation between the actual value and the target value. Suppose that the performance deviation follows a mixed distribution consisting of a normal distribution and a generalized Pareto distribution; The expectation-maximization algorithm is used to estimate the weights of the mixed distribution, the parameters of the normal distribution, and the parameters of the generalized Pareto distribution to form a mixed distribution model; Based on the mixed distribution model, the cumulative probability of performance deviation being lower than each preset management anchor threshold is calculated as the tail risk probability; The expected value of performance deviation is calculated as the conditional value of risk, given that the performance deviation is already below the management anchor threshold.

[0039] Furthermore, quantify the additional risk increment caused by each fundamental trigger, including: Based on the reasons in the root cause ranking list, the historical data is divided into two groups: one group where the reason is in an abnormal state and the other group where it is in a normal state. Calculate the tail risk probability based on the two sets of data respectively; The difference between the two sets of tail risk probabilities is taken as the additional risk increment caused by the fundamental cause.

[0040] Specifically, the purpose of constructing a mixed extreme value distribution model is to refine the statistical characteristics of performance deviation, paying particular attention to the probability and severity of its extreme adverse events. Performance deviation is defined as the relative deviation between the actual performance value of a project and the preset target value, usually expressed as a percentage, with negative values ​​representing unmet performance gaps.

[0041] Specifically, to simultaneously characterize both normal volatility and extreme risk, we assume that the performance deviation data follows a mixture distribution. This mixture distribution is composed of a normal distribution representing normal volatility and a generalized Pareto distribution (GPD) specifically characterizing extreme deviations. Let the mixture weights be... Then the probability density function of performance deviation It can be represented as: ,in, The mean is variance is The probability density function of the normal distribution, its parameters and It characterizes the typical fluctuation level and dispersion of performance deviation. Let the probability density function of the generalized Pareto distribution (GPD) be denoted by its shape parameter. Scale parameters and threshold parameters These factors collectively determine the shape and thickness of the extreme bias tail. The model takes historical and current performance bias data sequences as input and iteratively estimates the mixed weights using the Expectation-Maximization (EM) algorithm. Normal distribution parameters and and generalized Pareto distribution parameters , and This allows for probabilistic modeling of the overall data generation mechanism.

[0042] Based on a well-fitted mixture distribution model, risk indicators of concern to management can be calculated. Management typically presets multiple risk level anchor thresholds, such as... Corresponding to mild risk, This corresponds to significant risks. For each threshold... The cumulative probability of performance deviation falling below this threshold is calculated as the tail risk probability. This probability quantifies the likelihood of performance falling into the "failing" range. Further, the conditional value of risk is calculated. That is, when the performance deviation has been determined to be below the threshold. Expected deviation under the given conditions: ,in, This represents the mathematical expectation operator. This value indicates the average severity of loss that is expected when a risk event occurs, providing a quantitative basis for estimating risk reserves.

[0043] To establish the correlation between risk and root causes, the additional risk increment caused by each root cause was quantified. Based on the aforementioned generated ranking list of root causes, the primary causes were selected. The historical data is divided into two groups: one group contains data under abnormal conditions caused by the trigger, and the other group contains data under normal conditions. Based on these two groups of data, the tail risk probability is calculated using the same method described above, denoted as... and The difference between the two That is the trigger. In risk level This represents the additional risk increment caused by various management loopholes or external shocks. This increment visually reveals the marginal contribution of different management loopholes or external shocks to triggering extreme performance risks, enabling risk management to be more targeted.

[0044] Based on the root cause ranking list, tail risk probability and conditional risk value, and additional risk increment, a multi-objective optimization model is constructed to generate resource reallocation schemes, and the resource reallocation schemes are transformed into executable control instructions. Furthermore, a multi-objective optimization model is constructed to generate a resource reallocation scheme, including the following steps: The resource allocation for each budget item is used as the decision variable; The optimization objective is to maximize the total expected performance improvement of all projects and minimize the total extreme risks of all projects. The constraints are the total adjustable amount of each budget item and the basic resource requirements for project operation; Based on the optimization objectives and constraints, a multi-objective optimization function is constructed and solved to obtain the Pareto optimal solution as a resource reallocation scheme, and the expected performance improvement value corresponding to the scheme is recorded.

[0045] Specifically, the construction of a multi-objective optimization model aims to transform the diagnostic and risk information obtained from the aforementioned analysis into scientific and quantitative resource allocation decisions. This step, through the establishment of a mathematical model, seeks the optimal balance between simultaneously improving overall performance expectations and controlling systemic risks under the constraint of limited adjustable resources.

[0046] First, define the decision variables, which are the amount of resources to be additionally allocated to each budget item according to the plan. A common setup is that, for the , Each budget item has a decision variable that is a scalar. This represents the total amount of additional resources to be allocated to it. In a more refined model, the decision variables can be further refined into a two-dimensional form. , indicating from the first The budget item was adjusted to the first The specific amount of resources for each project.

[0047] Next, optimization objectives are constructed, setting two objectives that need to be optimized synergistically. The first objective is to maximize the overall expected performance improvement, based on the average treatment effect and intervention weight of the root causes of each project obtained from the aforementioned counterfactual attribution analysis. The second objective is to minimize the overall extreme risk, based on the conditional risk value of each project output by the risk quantification model. Alternatively, the second objective can be transformed into maximizing the return from maximizing the overall risk reduction. The resource performance elasticity coefficient for each project is obtained by training with historical data. Risk mitigation coefficient This allows us to quantify the two objectives mentioned above as functions of decision variables. For example, the project... The expected performance improvement can be expressed as Its extreme risk reduction return can be expressed as .

[0048] Based on this, a multi-objective optimization function is constructed. A linear weighted method is used to transform the bi-objective problem into a single-objective problem, the mathematical expression of which is as follows: ,in, This represents the total number of budget items participating in the optimization. For the project The resource performance elasticity coefficient is the expected performance improvement that can be achieved by adding one unit of resource to the project. (Symbol) For the project The risk mitigation coefficient is the expected reduction in extreme risk for each additional unit of resource added to the project. (Symbol) It is a weighting coefficient between 0 and 1, used to balance the trade-offs between the two management dimensions of performance improvement and risk control. This optimization function uses the decision variables of all projects... Elasticity coefficient and and weight As input, output a scalar optimization target value that integrates performance and risk considerations.

[0049] Next, constraints are set. These constraints typically include upper limits on the total amount of adjustable resources for each budget item, as well as minimum and maximum resource limits for each budget item to maintain basic operation or effectively absorb resources. For the use of two-dimensional decision variables... The model's subject quota constraint can be expressed as a constraint on each subject. The total resources allocated to all projects must not exceed their available quota. Project resource constraints are typically expressed as follows for each project. The total resources it obtained It must fall within the preset range Inside, among which As the lower limit of basic demand, To make effective use of the upper limit.

[0050] Finally, a linear programming or mixed-integer programming solver is invoked to solve the constrained optimization problem. The solution yields a set of resource allocation quantities, i.e., resource reallocation schemes. This scheme is Pareto optimal, meaning that no further optimization is possible without compromising any objective. The total expected performance improvement calculated under this scheme is recorded as a benchmark for subsequent tracking and control of the implementation of control measures. This optimization process combines managerial experience with the model's quantitative calculations to generate a data-driven, precise control scheme.

[0051] The system tracks the execution effect of control commands and performs feedback calibration on the parameters of dynamic cause-effect graphs, structural equation models, and mixed extreme value distribution models based on the execution effect.

[0052] Furthermore, the execution effect of control commands is tracked, and the parameters of the dynamic cause-effect graph, structural equation model, and mixed extreme value distribution model are calibrated based on the execution effect, including the following steps: Calculate the actual performance improvement within a preset time window after the execution of control instructions; Compare the actual degree of performance improvement with the expected performance improvement; When the actual improvement reaches the expected performance improvement value, the causal coefficient of the path related to this intervention in the dynamic causal diagram will be increased proportionally. If the actual improvement does not reach the expected improvement value, the relevant causal structure is rediscovered and reassessed. The current control order, the corresponding resource reallocation plan, and the actual degree of performance improvement will be used as new samples and added to the historical database to retrain the mixed extreme value distribution model.

[0053] Specifically, tracking the execution effect of control instructions and calibrating the model based on this feedback is a step towards self-optimization and continuous improvement, thus forming a complete intelligent closed loop of analysis, decision-making, execution and learning.

[0054] Specifically, after the resource reallocation plan is issued and executed as a control instruction, its workflow is not terminated. Instead, the operational status of relevant budget items under the new resource conditions is continuously monitored. After a preset evaluation time window, the actual performance improvement during that period is calculated. In some embodiments, the actual performance improvement... This is obtained by calculating the rate of change of key performance indicators before and after implementation.

[0055] Next, the degree of actual performance improvement will be assessed. The expected performance improvement value corresponding to this scheme as recorded by the multi-objective optimization model. A comparison is then made. The results of this comparison are used to verify the accuracy of previous causal diagnoses and risk predictions, and to drive iterative updates to the model.

[0056] If the actual improvement meets or exceeds expectations, it indicates that the intervention based on the current causal relationship is effective. At this point, the confidence in this successful experience will be increased. The weights or coefficients of the causal paths directly related to this intervention in the dynamic causal diagram will be increased proportionally. This increase can be performed using the following formula: ,in, The original causal coefficients in the structural equation model represent the causal path to be enhanced, and are quantitative estimates of the influence of the causal variable on the outcome variable. This is the preset learning rate, used to control the step size of model updates. (Symbol) The degree of actual observed performance improvement This represents the performance improvement value previously predicted by the model. The update mechanism takes the old coefficients, learning rate, and the difference between the actual and expected improvement as input, and outputs the updated causal coefficients. This enables reinforcement learning of effective causal relationships.

[0057] Conversely, if the actual improvement is significantly lower than expected, it suggests that the current causal structure may be biased or incomplete, failing to fully reflect the true management logic. In this case, a process for rediscovering and evaluating the relevant causal structure is triggered. New data containing the current intervention case is incorporated into the sliding time window, and the aforementioned causal discovery algorithm integrating domain knowledge is rerun, which may update the structure of the dynamic causal graph and re-estimate the structural equation model.

[0058] Finally, regardless of whether the intervention results meet expectations, the complete decision-making and implementation case, including the issued control instructions, the corresponding resource reallocation plan, and the ultimately observed degree of performance improvement, will be added to the historical database as a new data sample. The expanded historical data will be used periodically to retrain the mixed extreme value distribution model, thereby updating the estimates of tail risk probability and conditional value of risk, ensuring that the risk quantification model can reflect the latest operational status and risk characteristics in a timely manner. Through this feedback and calibration mechanism, the system possesses the ability to continuously learn and evolve from management practice.

[0059] Example 2: In a second embodiment of the present invention, the present invention provides a budget performance and fund utilization linkage analysis and closed-loop management system, such as... Figure 2 As shown, it includes the following modules: Acquisition and processing module: used to acquire multi-source heterogeneous time series data during the execution of budget projects, and to fuse and perform feature engineering on the multi-source heterogeneous time series data to form a standardized multi-dimensional time series feature matrix; Constructing a reflection module: This module is used to construct a dynamic causal graph based on a multi-dimensional time-series feature matrix and integrate prior knowledge of the domain, and to estimate a structural equation model reflecting the causal relationship between funding variables and performance variables based on the dynamic causal graph. The reasoning and calculation module is used to perform counterfactual reasoning and attribution analysis on the detected performance anomalies based on structural equation modeling, obtain performance deviation data, calculate the average treatment effect of each candidate cause, and generate a ranking list of root causes by combining interventionist weights. The computational quantification module is used to construct a mixed extreme value distribution model based on historical and current performance deviation data, calculate the tail risk probability and conditional risk value under a preset management anchor threshold, and quantify the additional risk increment caused by each fundamental cause. The generation and transformation module is used to build a multi-objective optimization model based on the root cause ranking list, tail risk probability and conditional risk value, and additional risk increment to generate resource reallocation schemes, and to transform resource reallocation schemes into executable control instructions. Tracking and Optimization Module: Used to track the execution effect of control commands and to perform feedback calibration on the parameters of dynamic cause-effect graphs, structural equation models and mixed extreme value distribution models based on the execution effect.

[0060] A finance department, responsible for managing budget projects across multiple livelihood-related sectors including education, healthcare, and infrastructure, faces numerous challenges in practice. The allocation of funds is poorly linked to performance output, resulting in some projects experiencing idle and wasted funds, while others are hampered by insufficient resources. Risk warnings lack timeliness; traditional management models rely on manual verification, making it difficult to quickly identify potential problems such as irregular fund usage and lagging performance targets during project execution. When performance anomalies occur, the root causes are often unclear, making it difficult to accurately pinpoint the problem's origin and leading to a lack of targeted rectification. Consequently, fiscal resources fail to fully realize their potential value and are unable to efficiently respond to the public's actual needs for livelihood services. To address these issues, the budget performance and fund usage linkage analysis and closed-loop management system provided by this invention was adopted, with the architecture as follows: Figure 2 As shown. The specific implementation process of this system is as follows: First, the acquisition and processing module comprehensively collects multi-source heterogeneous time-series data such as project budget indicators, fund allocation details, performance target filing forms, and external economic data. Through cleaning, standardization, and feature engineering, a standardized multi-dimensional time-series feature matrix is ​​formed to eliminate data heterogeneity interference and provide a high-quality data foundation for subsequent analysis. Subsequently, a module is constructed to integrate prior knowledge in the fiscal field, a dynamic causal graph is built, and the causal relationship between funds and performance variables is accurately characterized. Then, a structural equation model is estimated to quantify the influence between variables and provide a scientific basis for anomaly attribution. Next, the reasoning and calculation module conducts counterfactual reasoning and attribution analysis through structural equation modeling for the detected performance anomalies, calculates the average treatment effect of each candidate cause, and generates a root cause ranking list by combining interventionist weights to clarify the core issues. Then, the calculation and quantification module constructs a hybrid extreme value distribution model based on historical and current performance deviation data, calculates the tail risk probability and conditional risk value under the preset threshold, and quantifies the additional risk increment of each fundamental cause to achieve accurate risk quantification. Then, by generating a conversion module, the list of causes and risk data are integrated to build a multi-objective optimization model, generate a Pareto optimal resource reallocation plan, and convert it into executable control instructions to guide the precise allocation of funds; Finally, the tracking and optimization module tracks the execution effect of instructions in real time, compares the actual performance improvement with the expected improvement, and provides feedback to calibrate the parameters of the dynamic cause-effect graph, structural equation model and mixed extreme value distribution model, forming a closed loop of "analysis-decision-execution-optimization" to improve the efficiency of fiscal resource utilization and risk management capabilities.

[0061] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for linking budget performance with fund utilization and implementing closed-loop management, characterized in that... Includes the following steps: Acquire multi-source heterogeneous time-series data during the execution of budget projects, and fuse and perform feature engineering on the multi-source heterogeneous time-series data to form a standardized multi-dimensional time-series feature matrix; Based on the multidimensional time-series feature matrix, a dynamic causal graph is constructed by integrating prior knowledge of the domain, and a structural equation model reflecting the causal relationship between funding variables and performance variables is estimated based on the dynamic causal graph. Based on the structural equation model, counterfactual reasoning and attribution analysis are performed on the detected performance anomalies to obtain performance deviation data, calculate the average treatment effect of each candidate cause, and generate a root cause ranking list by combining interventionist weights. Based on historical and current performance deviation data, a hybrid extreme value distribution model is constructed to calculate the tail risk probability and conditional risk value under a preset management anchor threshold, and to quantify the additional risk increment caused by each fundamental cause. Based on the fundamental cause ranking list, the tail risk probability and conditional risk value, and the additional risk increment, a multi-objective optimization model is constructed to generate a resource reallocation scheme, and the resource reallocation scheme is transformed into an executable control instruction. The execution effect of the control instructions is tracked, and the parameters of the dynamic causal graph, the structural equation model, and the mixed extreme value distribution model are calibrated based on the execution effect.

2. The method for linking budget performance and fund utilization analysis and closed-loop management according to claim 1, characterized in that: The formation of a standardized multidimensional temporal feature matrix includes the following steps: Real-time acquisition of cash flow data, performance indicator data, and environmental context data, and alignment and interpolation at a uniform time granularity; Calculate budget execution rate, expenditure volatility and payment concentration characteristics from cash flow data; calculate achievement speed and quality stability characteristics from performance indicator data; and extract policy impact factors and market volatility index characteristics from environmental context data through natural language processing and numerical calculation. All extracted features are organized according to time series to form a multidimensional time series feature matrix with uniform dimensions.

3. The method for linking budget performance and fund utilization analysis and closed-loop management according to claim 1, characterized in that: The construction of a dynamic causal graph based on the multidimensional temporal feature matrix and incorporating prior domain knowledge includes the following steps: The prior knowledge in the field of budget management is formalized into time-irreversible constraints, rigid budget transmission constraints, and lagging institutional and process constraints, thus forming prior knowledge constraints. Within the sliding time window, based on the prior knowledge constraints, an improved causal inference algorithm is executed on the multidimensional time series feature matrix to perform conditional independence testing and causal direction determination. The output part is a directed acyclic graph as the causal structure of the current window, and the stability of the structure is evaluated by calculating the edit distance of the causal graphs of adjacent windows. The structure is updated when it changes, forming a dynamic causal graph.

4. The method for linking budget performance and fund utilization analysis and closed-loop management according to claim 1, characterized in that: The structural equation model that reflects the causal relationship between funding variables and performance variables based on the dynamic causal graph includes the following steps: For each variable in the dynamic causal graph, the structural equation coefficients are estimated using a linear regression method with signed constraints, with its parent node as the independent variable. The structural equation coefficients are used to construct linear equations for each variable, forming a structural equation model that characterizes the quantitative causal relationship between the variables.

5. The method for linking budget performance and fund utilization analysis and closed-loop management according to claim 1, characterized in that: The process of performing counterfactual reasoning and attribution analysis on the detected performance anomalies to obtain performance deviation data and calculate the average treatment effect of each candidate cause includes the following steps: Locate all candidate cause nodes leading to abnormal performance variables in the dynamic cause-effect graph; For each candidate cause, within the framework of the structural equation model, an intervention is performed to set the cause variable to its planned value or normal state value; Calculate the counterfactual performance expectation after intervention, and use the difference between the observed performance and the counterfactual performance expectation as the average treatment effect of the candidate cause; The difference between the observed performance and the expected counterfactual performance is recorded as the performance deviation data for the current monitored event.

6. The method for linking budget performance and fund utilization analysis and closed-loop management according to claim 1, characterized in that: The generation of the root cause ranking list by combining the interventionability weights includes: Assign a weight coefficient representing the manageability of each candidate cause; The absolute value of the average treatment effect of each candidate cause after time-adjusted is normalized to obtain the original contribution. The original contribution score is multiplied by the corresponding interventionist weight coefficient and then normalized twice to obtain the ranking score. Candidate causes are sorted in descending order of their ranking scores to generate a root cause ranking list.

7. The method for linking budget performance and fund utilization analysis and closed-loop management according to claim 1, characterized in that: The construction of the hybrid extreme value distribution model, and the calculation of the tail risk probability and conditional value of risk under the preset management anchor threshold, includes the following steps: Performance deviation is defined as the relative deviation between the actual value and the target value. Suppose that the performance deviation follows a mixed distribution consisting of a normal distribution and a generalized Pareto distribution; The weights, normal distribution parameters, and generalized Pareto distribution parameters of the mixed distribution are estimated using the expectation-maximization algorithm to form a mixed distribution model; Based on the mixed distribution model, the cumulative probability of performance deviation being lower than each preset management anchor threshold is calculated as the tail risk probability; The expected value of performance deviation is calculated as the conditional value of risk, given that the performance deviation is already below the management anchor threshold.

8. The method for linking budget performance and fund utilization analysis and closed-loop management according to claim 1, characterized in that: The quantification of the additional risk increment caused by each fundamental cause includes: Based on the reasons in the root cause ranking list, the historical data is divided into two groups: one group where the reason is in an abnormal state and the other group where it is in a normal state. The tail risk probability is calculated based on the two sets of data respectively; The difference between the two sets of tail risk probabilities is taken as the additional risk increment caused by the fundamental cause.

9. The method for linking budget performance and fund utilization analysis and closed-loop management according to claim 1, characterized in that: The process of constructing a multi-objective optimization model to generate a resource reallocation scheme includes the following steps: The resource allocation for each budget item is used as the decision variable; The optimization objective is to maximize the total expected performance improvement of all projects and minimize the total extreme risks of all projects. The constraints are the total adjustable amount of each budget item and the basic resource requirements for project operation; Based on the optimization objectives and constraints, a multi-objective optimization function is constructed and solved to obtain the Pareto optimal solution as a resource reallocation scheme, and the expected performance improvement value corresponding to the scheme is recorded.

10. The method for linking budget performance and fund utilization analysis and closed-loop management according to claim 1, characterized in that: The process of tracking the execution effect of the control commands and performing feedback calibration on the parameters of the dynamic causal graph, the structural equation model, and the mixed extreme value distribution model based on the execution effect includes the following steps: Calculate the actual performance improvement within a preset time window after the execution of control instructions; Compare the actual degree of performance improvement with the expected performance improvement value; When the actual improvement reaches the expected performance improvement value, the causal coefficient of the path related to this intervention in the dynamic causal diagram will be increased proportionally. If the actual improvement does not reach the expected improvement value, the relevant causal structure is rediscovered and reassessed. The current control order, the corresponding resource reallocation plan, and the actual degree of performance improvement are used as new samples and added to the historical database to retrain the hybrid extreme value distribution model.

Images