Method and system for generating adaptive intervention strategies based on scm-based volatility attribution

Abstract

The application particularly relates to a SCM-based fluctuation attribution and adaptive intervention strategy generation method and system. The method comprises the following steps: acquiring multi-source macroeconomic time series data of a target region; constructing a differentiable structural causal model; monitoring a target economic indicator in real time; if the target economic indicator deviates from an expected trajectory, performing multivariate counterfactual disturbance analysis through the differentiable structural causal model, identifying core driving factors and generating a structural bias vector; generating a combined intervention strategy set based on the core driving factors, the structural bias vector and a pre-constructed intervention anchor knowledge graph; substituting the current economic state, the combined intervention strategy set and the structural bias vector into the differentiable structural causal model to generate a multi-dimensional combined intervention strategy and its time sequence execution parameters; and outputting intervention effect prediction and expected causal chain through counterfactual simulation verification. The method realizes explicit modeling of hidden variables, improves attribution accuracy and ensures causal consistency and interpretability of the strategy.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and macroeconomic decision-making technology, specifically providing a method and system for generating volatility attribution and adaptive intervention strategies based on SCM. Background Technology

[0002] Accurate attribution and effective intervention of macroeconomic fluctuations (such as economic downturns or overheating) are core challenges in economic research and policymaking. Current technologies suffer from the following main shortcomings: Existing methods for analyzing economic fluctuations are generally 1. Existing economic fluctuation analysis methods are difficult to distinguish the direction of cause and effect and cannot handle the latent variables that are common in the macroeconomic system (such as market expectations and policy uncertainty). Ignoring latent variables will lead to serious bias in the estimation of causal networks, which in turn will affect the accuracy of root cause identification. 2. Current policy-making relies heavily on expert experience and uses trial and error or macroeconomic econometric models for policy simulation, which is inefficient, subjective, and difficult to deal with complex economic systems. 3. Existing methods generally treat causal discovery and policy optimization as two independent stages, resulting in unsatisfactory causal consistency and sample efficiency of the policy. 4. Existing intervention strategies cannot generate coordinated intervention strategies by combining the policy effects of causal transmission in real economic decision-making.

[0003] Therefore, there is a need in this field for a method and system for generating fluctuation attribution and adaptive intervention strategies based on SCM that can integrate causal modeling and deep reinforcement learning to solve the above problems and achieve accurate attribution of economic fluctuations and generation of adaptive intervention strategies. Summary of the Invention

[0004] Based on this, the present invention proposes a method and system for generating volatility attribution and adaptive intervention strategies based on SCM, which solves the technical problems of existing economic volatility analysis methods, such as limited causal inference capabilities, reliance on human experience in intervention strategy generation, separation of strategy optimization and causal model, and inability to generate multi-tool synergistic intervention strategies for real economic decisions.

[0005] According to a first aspect of the present invention, a method for generating volatility attribution and adaptive intervention strategies based on SCM is provided, the method comprising the following steps: Acquire multi-source macroeconomic time-series data for the target region; Based on the aforementioned multi-source macroeconomic time-series data, a differentiable causal model is constructed. Real-time monitoring to see if target economic indicators deviate from the preset trajectory: If the target economic indicator deviates from the expected trajectory, a multivariate counterfactual disturbance analysis is performed based on the differentiable structural causal model to identify the core driving factors and generate a structural deviation vector. Based on the core driving factors, the structural deviation vector, and the pre-built intervention anchor knowledge graph, a set of combined intervention strategies is generated. The set of combined intervention strategies includes at least two candidate intervention actions of different action types and the time-series execution parameters corresponding to each candidate intervention action. Substitute the current economic state, the set of combined intervention strategies, and the structural deviation vector into the differentiable structural causal model to generate a multi-dimensional combined intervention strategy and the time-series execution parameters of the multi-dimensional combined intervention strategy. The multi-dimensional combined intervention strategy is verified by counterfactual simulation, and the predicted intervention effect and expected causal chain are output.

[0006] According to an embodiment of the present invention, constructing a differentiable causal model based on the multi-source macroeconomic time-series data includes: Based on the aforementioned multi-source macroeconomic time-series data, a dynamic structural causal model is constructed, wherein the dynamic structural causal model includes observable variables and latent variable nodes; The dynamic structural causal model is trained to obtain a differentiable structural causal model.

[0007] According to an embodiment of the present invention, the multi-source macroeconomic time-series data includes at least: regional GDP, industrial output value, consumption indicators, investment indicators, import and export indicators, money supply, price index, and policy text quantitative index.

[0008] According to an embodiment of the present invention, the step of performing multivariate counterfactual perturbation analysis based on the differentiable structural causal model to identify core driving factors and generate a structural bias vector includes: Using the target economic indicator as the target variable, obtain the set of parent nodes of the target variable in the differentiable causal model. For each parent node in the set of parent nodes, fix the other parent nodes as the current true observation value, and replace the parent node with several counterfactual values ​​under different quantiles; Based on the differentiable structural causal model, the do operator intervention is performed to identify the core driving factors and generate a structural deviation vector.

[0009] According to an embodiment of the present invention, the step of performing do operator intervention based on the differentiable structural causal model to identify core driving factors and generate a structural deviation vector includes: Based on the replaced parent node, the do operator intervention is performed based on the differentiable causal model to obtain the expected value of the target variable under each counterfactual value intervention; The average of the absolute differences between the expected value of the target variable under all quantiles and the expected value of the target variable when all parent nodes take the current true value is taken as the structural contribution of the parent node. The top few factors with the largest structural contributions are identified as the core driving factors; For each core driving factor, the mean of the historical stationary period of the parent node of the core driving factor is taken as the stationary state value. Based on the differentiable causal model, the counterfactual expected value of the core driving factor when the parent node takes the steady state value is determined, and the counterfactual expected value is used as the causal benchmark value of the core driving factor. The difference between the current value of each core driving factor and the causal baseline value of the core driving factor is divided by the standard deviation of the core driving factor during historical stationary periods to obtain the standardized deviation of each core driving factor. All standardized deviations are combined as the structural deviation vector.

[0010] According to an embodiment of the present invention, before generating a set of combined intervention strategies based on the core driving factors, the structural deviation vector, and the pre-constructed intervention anchor knowledge graph, the method further includes: Bayesian posterior calculations are performed on differentiable structural causal models to generate several posterior causal graph samples and corresponding structural equation parameter samples. Based on each posterior causal graph sample, determine the average causal effect of each candidate core driver on the target variable and the posterior distribution of each candidate core driver. Based on the average causal effect of each candidate core driver factor on the target variable and the posterior distribution of each candidate core driver factor, the causal effect confidence index of each candidate core driver factor is determined. The causal effect confidence index of each candidate core driver factor is used to screen for robust core drivers factors and factors to be observed.

[0011] According to an embodiment of the present invention, generating a set of combined intervention strategies based on the core driving factors, the structural deviation vector, and a pre-constructed intervention anchor knowledge graph includes: Based on the robust core driving factors, the pre-constructed intervention anchor knowledge graph is matched to obtain all influencing parameters; Based on all influencing parameters and the structural deviation vector, several candidate intervention actions and the corresponding time-series execution parameters for each candidate intervention action are generated. Based on several candidate intervention actions and the timing execution parameters corresponding to each candidate intervention action, a set of combined intervention strategies is constructed.

[0012] According to a second aspect of the present invention, a fluctuation attribution and adaptive intervention strategy generation system based on SCM is provided, for implementing the steps of any of the above-described methods, characterized in that the system comprises: The data acquisition module is used to acquire multi-source macroeconomic time-series data and preprocess the multi-source macroeconomic time-series data to obtain multi-source macroeconomic time-series data of the target region. The differentiable causal discovery module is used to obtain a differentiable structural causal model based on multi-source macroeconomic time-series data of the target region. The detection module is used to monitor the fluctuations of the target economic indicator in real time, and based on the fluctuations of the target indicator, detect whether the target economic indicator deviates from the preset trajectory. The causal attribution module is used to perform multivariate counterfactual perturbation analysis based on the differentiable structural causal model when the target economic indicator deviates from the expected trajectory, identify the core driving factors and generate a structural deviation vector. The combined intervention strategy generation module is used to generate a set of combined intervention strategies based on the core driving factors, the structural deviation vector, and the pre-constructed intervention anchor knowledge graph. The intervention anchor knowledge graph module is used to store and manage the causal relationship knowledge between economic fluctuation types, driving factor categories and policy tools, so as to facilitate retrieval and reasoning based on the semantics of candidate intervention anchors; The combined intervention optimization module is used to simultaneously solve the joint optimal solution of three dimensions—intervention tool selection, intervention intensity, and execution order—based on the current economic state, the set of combined intervention strategies, and the structural deviation vector, through the Jacobian matrix of the differentiable structural causal model, and generate a multi-dimensional combined intervention strategy and the time-series execution parameters of the multi-dimensional combined intervention strategy. The counterfactual simulation and interpretable output module is used to perform multi-scenario counterfactual simulations on the multi-dimensional combined intervention strategy, and output the intervention effect prediction range, causal chain visualization and decision suggestions.

[0013] According to an embodiment of the present invention, the system further includes a robust screening module, the robust screening module comprising: The Bayesian posterior computation unit is used to perform weighted resampling on the original dataset in the differentiable structural causal model to obtain several weighted sampled datasets, and to re-estimate the causal graph structure and structural equation parameters on each weighted sampled dataset to generate several posterior causal graph samples and corresponding structural equation parameter samples. The average causal unit is used to construct a set of candidate core driving factors and a posterior differentiable causal model based on each posterior causal graph sample, and to determine the average causal effect of each candidate core driving factor on the target variable and the posterior distribution of each candidate core driving factor based on the set of candidate core driving factors and the posterior distribution of each candidate core driving factor. The confidence index calculation unit is used to determine the causal effect confidence index of each candidate core driver factor based on the average causal effect of each candidate core driver factor on the target variable and the posterior distribution of each candidate core driver factor. The driving factor screening unit is used to screen the causal effect confidence index of each candidate core driving factor to obtain robust core driving factors and the factors to be observed.

[0014] According to an embodiment of the present invention, the combined intervention strategy generation module includes: The knowledge graph matching unit is used to determine the corresponding economic variable node of the robust core driving factor in the knowledge graph based on the robust core driving factor, and to perform a reverse breadth-first search in the knowledge graph based on each economic variable node as the target node to determine the policy tool node of each robust core driving factor and obtain all the influencing parameters. An intervention action generation unit is used to generate several candidate intervention actions and the corresponding time-series execution parameters for each candidate intervention action based on all influencing parameters and the structural deviation vector. The combined strategy generation unit is used to determine the priority of each candidate intervention action based on several candidate intervention actions and the timing execution parameters corresponding to each candidate intervention action, and to construct a combined intervention strategy set according to the priority of each candidate intervention action.

[0015] As can be seen from the above technical solutions, the fluctuation attribution and adaptive intervention strategy generation method and system based on SCM provided by the present invention have the following beneficial effects: By constructing a differentiable structural causal model based on multi-source macroeconomic time-series data of the target region, the unobservable factors of the target region are made explicit, thereby reducing the error rate of causal network estimation. Real-time monitoring of target economic indicators enables multivariate counterfactual disturbance analysis to identify core driving factors and generate structural deviation vectors when deviations occur, allowing for timely differentiation between structural deviations and cyclical fluctuations. Based on these core driving factors, structural deviation vectors, and a pre-constructed intervention anchor knowledge graph, a set of combined intervention strategies is generated, enabling the automatic generation of intervention strategy sets across different dimensions. By substituting the aforementioned data into the differentiable structural causal model, multi-dimensional combined intervention strategies and their time-series execution parameters are generated, achieving one-time optimization of multi-dimensional joint variables. Its time-series awareness makes strategy generation more closely resemble real policy-making scenarios. Counterfactual simulation verification of the multi-dimensional combined intervention strategies outputs predicted intervention effects and expected causal chains, enabling reverse verification of the multi-dimensional combined intervention strategies. This facilitates corporate decision-makers' understanding of the principles and effects of policies, thereby ensuring the causal consistency of the strategies. Attached Figure Description

[0016] The disclosure of this invention will become more readily understood with reference to the accompanying drawings. It will be readily understood by those skilled in the art that these drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. Furthermore, similar numbers in the drawings are used to denote similar components, wherein: Figure 1 This is a schematic diagram of the main process of a fluctuation attribution and adaptive intervention strategy generation method based on SCM according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the counterfactual attribution and strategy optimization core process of a fluctuation attribution and adaptive intervention strategy generation method based on SCM according to an embodiment of the present invention. Figure 3 This is a schematic diagram of a differentiable causal model of a fluctuation attribution and adaptive intervention strategy generation method based on SCM according to an embodiment of the present invention. Figure 4 This is a schematic diagram of the main structure of a fluctuation attribution and adaptive intervention strategy generation system based on SCM according to an embodiment of the present invention. Detailed Implementation

[0017] Some embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0018] In the description of this invention, "module" and "processor" can include hardware, software, or a combination of both. A module can include hardware circuitry, various suitable sensors, communication ports, memory, and may also include software components, such as program code, or a combination of software and hardware. A processor can be a central processing unit, microprocessor, image processor, digital signal processor, or any other suitable processor. The processor has data and / or signal processing capabilities. The processor can be implemented in software, in hardware, or a combination of both. Non-transitory computer-readable storage media includes any suitable medium capable of storing program code, such as magnetic disks, hard disks, optical disks, flash memory, read-only memory, random access memory, etc. The term "A and / or B" means all possible combinations of A and B, such as only A, only B, or A and B. The terms "at least one A or B" or "at least one of A and B" have a similar meaning to "A and / or B" and can include only A, only B, or A and B. The singular terms "a" or "this" can also include plural forms.

[0019] See appendix Figures 1 to 3 , Figure 1 This is a schematic diagram of the main process of a fluctuation attribution and adaptive intervention strategy generation method based on SCM according to an embodiment of the present invention. Figure 1 As shown in the figure, a method for generating fluctuation attribution and adaptive intervention strategies based on SCM in an embodiment of the present invention includes the following steps: Step S1: Obtain multi-source macroeconomic time-series data for the target region; In some embodiments, acquiring multi-source macroeconomic time-series data for the target region includes: Define the target area; Collect multi-source macroeconomic time-series raw data for the target region; The multi-source macroeconomic time series raw data is preprocessed to obtain preprocessed multi-source macroeconomic time series raw data as the multi-source macroeconomic time series data for the target region.

[0020] In some embodiments, the multi-source macroeconomic time-series raw data originates from national or regional statistical departments, central banks, the General Administration of Customs, and other official economic data release platforms. Natural language processing is performed on the raw data, that is, policy texts such as government work reports, central bank monetary policy implementation reports, and fiscal budget reports are quantified to generate policy text quantitative indices. These policy text quantitative indices include at least the following dimensions: policy direction coding category (which can be easing / tightening / neutral), policy strength score, and policy implementation time window.

[0021] In some embodiments, the data time span of the acquired multi-source macroeconomic time series raw data covers at least 10 to 20 quarters to ensure the statistical power of causal discovery.

[0022] In some embodiments, the multi-source macroeconomic time-series data includes at least: regional GDP, industrial output, consumption indicators, investment indicators, import and export indicators, money supply, price indices, and quantitative indices based on policy documents.

[0023] Step S2: Based on the multi-source macroeconomic time-series data, construct a differentiable causal model; In some embodiments, constructing a differentiable structural causal model based on the multi-source macroeconomic time-series data includes: Based on the aforementioned multi-source macroeconomic time-series data, a time-series causal discovery algorithm incorporating latent variables is used to construct a dynamic structural causal model, which includes observable variables and latent variable nodes. The dynamic structural causal model is trained to obtain a differentiable structural causal model.

[0024] In some embodiments, the step of constructing a dynamic structural causal model specifically includes: The dynamic structural causal model adopts a neural structural causal model framework, in which the structural equation of each variable is parameterized by a differentiable neural network, and the causal structure between variables is learned through the NOTEARS algorithm. This algorithm transforms the discrete directed acyclic graph search problem into a continuous optimization problem by defining a weighted adjacency matrix. The optimization objective is to minimize the sum of the squared loss of the prediction error and the L1 sparse regularization term. The constraint condition is a directed acyclic graph constraint. Furthermore, for time series data, an additional lag constraint is added on the basis of the above optimization objective, and the lag order of the influence of each variable on the target variable is determined by cross-correlation analysis.

[0025] The differentiable structural causal model adopts a neural structural causal model framework, in which the structural equation of each variable is parameterized by a differentiable neural network, and the causal structure between variables is jointly optimized by a differentiable graph structure learning method and domain knowledge constraints, enabling the model to support end-to-end differentiable inference for any counterfactual scenario.

[0026] In some embodiments, training the dynamic structural causal model to obtain a differentiable structural causal model includes: The structural equation for each variable is parameterized into a differentiable neural network. The value of each endogenous variable is obtained by mapping the neural network to the value of its parent node, and exogenous noise variables are added. All the above data were jointly trained using a variational inference framework. The loss function consisted of three weighted components: a reconstruction likelihood term, which is the expected log-likelihood of reconstructing the observable variables by sampling from the posterior distribution of the latent variables; a KL divergence regularization term for the latent variable prior, which constrains the difference between the posterior distribution of the latent variables and the standard Gaussian prior; and a DAG constraint term, which ensures that the learned causal graph satisfies the directed acyclic graph property. The training employed the Adam optimizer with a learning rate of 0.001 until the loss function converged. This enables the differentiable causal model obtained after training to perform forward inference prediction and to intervene in counterfactual inference and do operator of any node—that is, to artificially set a certain variable to take a specific value (i.e., intervention) and propagate the corresponding changes of other variables, and this propagation process is completely differentiable.

[0027] In some embodiments, the step of constructing a differentiable causal model specifically includes: Let the set of observable variables include regional GDP, output value of various industries, consumption indicators, investment indicators, import and export indicators, money supply, price index, and quantitative indices from policy documents (no fewer than 8 variables), and let the set of latent variables include at least 2 latent variable nodes, representing economic expectation indicators and market uncertainty indicators, respectively: That is, let there be a set of observable variables. Where d is the number of variables, including regional GDP, industrial output, consumption indicators, investment indicators, import and export indicators, money supply, price index and policy text quantitative index.

[0028] Latent variable set , where m is the number of latent variables, including at least economic expectation indicators and policy uncertainty indicators.

[0029] The NOTEARS algorithm is used to learn the causal structure between variables, where the weighted adjacency matrix is ​​defined as follows: The optimization objective is:

[0030] Where n is the sample size. Let be the regularization coefficient, and the constraint condition is:

[0031] The structural equation for each variable is parameterized into a neural network, represented as follows:

[0032] in, for The set of parent nodes can contain observable variables and latent variables; For exogenous noise variables; It is a differentiable function; These are the network parameters to be learned.

[0033] By training all the modules mentioned above together, a differentiable causal model is obtained.

[0034] Step S3: Monitor in real time whether the target economic indicators deviate from the preset trajectory; In some embodiments, the specific steps for determining whether the target economic indicator deviates from the preset trajectory include: Set a target indicator Y (which could be the GDP growth rate) and determine its expected trajectory; The expected trajectory can be determined in various ways, including but not limited to: the historical average and its confidence interval, the predicted value of the time series model (which can be exponentially smoothed), and the counterfactual steady-state value of the structural causal model under the assumption of no external shock. Furthermore, the preset threshold can be an absolute deviation (which can be 1.5 percentage points lower than expected) or a relative deviation (which can also be 1.5 times lower than expected).

[0035] Step S4: If the target economic indicator deviates from the expected trajectory, perform multivariate counterfactual perturbation analysis based on the differentiable structural causal model to identify the core driving factors and generate a structural deviation vector; In some embodiments, performing multivariate counterfactual perturbation analysis based on the differentiable structural causal model to identify core driving factors and generate a structural bias vector includes: Using the target economic indicator as the target variable, obtain the set of parent nodes of the target variable in the differentiable causal model. For each parent node in the set of parent nodes, fix the other parent nodes as the current true observation value, and replace the parent node with several counterfactual values ​​under different quantiles; Based on the differentiable structural causal model, the do operator intervention is performed to identify the core driving factors and generate a structural deviation vector.

[0036] In some embodiments, the step of performing do operator intervention based on the differentiable structural causal model to identify core driving factors and generate a structural deviation vector includes: Based on the replaced parent node, the do operator intervention is performed based on the differentiable causal model to obtain the expected value of the target variable under each counterfactual value intervention; The average of the absolute differences between the expected value of the target variable under all quantiles and the expected value of the target variable when all parent nodes take the current true value is taken as the structural contribution of the parent node. The top few factors with the largest structural contributions are identified as the core driving factors; For each core driving factor, the mean of the historical stationary period of the parent node of the core driving factor is taken as the stationary state value. Based on the differentiable causal model, the counterfactual expected value of the core driving factor when the parent node takes the steady state value is determined, and the counterfactual expected value is used as the causal benchmark value of the core driving factor. The difference between the current value of each core driving factor and the causal baseline value of the core driving factor is divided by the standard deviation of the core driving factor during historical stationary periods to obtain the standardized deviation of each core driving factor. All standardized deviations are combined as the structural deviation vector.

[0037] In some embodiments, the specific steps of step S4 include: Let the set of parent nodes of the target variable Y in a differentiable causal model be . ; For each parent node Perform the following counterfactual experiment: That is, keep the values ​​of other parent nodes unchanged from the current actual observation values, Replace them sequentially with counterfactual values ​​at k different quantiles (which can be the 10th, 25th, 50th, 75th, and 90th percentiles); Execute do( using a differentiable structural causal model) =counterfactual value) intervention operation, calculate the expected value of Y under this counterfactual intervention; The difference between the expected value and the current true value of the target variable Y under counterfactual intervention is taken as the effect size of this counterfactual experiment, and the average of the effect sizes at all k quantiles is taken as... The structural contribution to Y fluctuations makes it possible to consider both It considers the marginal impact at different value levels while avoiding the randomness caused by using only a single variable.

[0038] That is, for the target variable Each parent node Fix other parent nodes in the actual observation values; Will Replace with counterfactual values ​​at k different quantiles, and run do The expected change of Y is calculated after calculating the counterfactual value, and the average value is taken as the mean. Structural contribution; The top K factors with the largest contributions are identified as the core driving factors; The structural bias vector is defined as the difference between the current value and the causal baseline value, wherein the causal baseline value is generated counterfactually by the differentiable structural causal model under the assumption of no external intervention.

[0039] In Embodiment 1, the specific steps of step S4 include: Real-time monitoring of target economic indicators This refers to the GDP growth rate, with a preset expected trajectory threshold. When detected... When the deviation from the expected trajectory exceeds a preset threshold: Let the target variable be... The set of parent nodes is For each parent node Fix other parent nodes as the current true observation values ,Will Replace with the first Counterfactual value at the quantile Calculate using structural causal model The expected change. The formula for calculating the structural contribution is:

[0040] in, The number of quantiles is taken here. =5, This indicates an intervention operation.

[0041] Normalize the structural contribution:

[0042] The top contributors to normalization These factors were identified as core driving factors.

[0043] For each core driving factor Calculate its causal baseline value:

[0044] in, This represents the stable state value of the parent node under the assumption of no external intervention. The structural deviation vector is defined as:

[0045] in, for The standard deviation during a period of historical stability.

[0046] In the above embodiments, after calculating the contribution of each parent node, the top K factors with the largest contribution are identified as core driving factors. Furthermore, the setting of the structural deviation vector facilitates the distinction between structural deviation (i.e., fundamental deviation caused by changes in causal mechanisms) and periodic fluctuations (i.e., short-term fluctuations caused by temporary shocks), thereby providing more accurate input for the formulation of subsequent intervention strategies.

[0047] Step S5: Based on the core driving factors, the structural deviation vector, and the pre-constructed intervention anchor knowledge graph, generate a set of combined intervention strategies. The set of combined intervention strategies includes candidate intervention actions of at least two different action types and the time-series execution parameters corresponding to each candidate intervention action. The action types include at least tax adjustments, subsidy disbursements, and industry guidance measures. The time-series execution parameters include synchronous execution, sequential execution, or staggered execution. In some embodiments, before generating a set of combined intervention strategies based on the core driving factors, the structural bias vector, and the pre-constructed intervention anchor knowledge graph, the method further includes: Perform Bayesian posterior calculations on the differentiable structural causal model to generate several posterior causal graph samples and corresponding structural equation parameter samples, wherein the number of posterior causal graph samples is no less than 100. Based on each posterior causal graph sample, determine the average causal effect of each candidate core driver on the target variable and the posterior distribution of each candidate core driver. Based on the average causal effect of each candidate core driver factor on the target variable and the posterior distribution of each candidate core driver factor, the causal effect confidence index of each candidate core driver factor is determined. The causal effect confidence index of each candidate core driver factor is used to screen for robust core drivers factors and factors to be observed.

[0048] In some embodiments, the causal effect confidence index is expressed as:

[0049] in, For the first The average causal effect of each candidate driver factor This is the preset minimum economic significance threshold.

[0050] In some embodiments, the screening of the causal effect confidence index of candidate core driving factors is as follows: to retain The factors are taken as robust core driving factors, and the remaining factors are taken as factors to be observed.

[0051] In some embodiments, generating a set of combined intervention strategies based on the core driving factors, the structural bias vector, and a pre-constructed intervention anchor knowledge graph includes: Based on the robust core driving factors, the pre-constructed intervention anchor knowledge graph is matched to obtain all influencing parameters; Based on all influencing parameters and the aforementioned structural bias vector, several candidate intervention actions and corresponding time-series execution parameters for each candidate intervention action are generated. The structured information included in each candidate intervention action includes at least the following: action type (specifically, tax adjustment, subsidy distribution, industrial guidance, and monetary adjustment), target (specifically, specific industry, enterprise type, or economic variable), intensity range (specifically, feasible intervals for continuous parameters or enumeration of discrete options), and expected causal path (specifically, the causal chain from the intervention node through the mediating variable to the target variable, with an estimated causal effect). Based on several candidate intervention actions and the timing execution parameters corresponding to each candidate intervention action, a set of combined intervention strategies is constructed.

[0052] In some embodiments, the pre-built intervention anchor knowledge graph can be a causal perception graph structure knowledge base, wherein nodes include two types of entities: economic variables (which may be GDP, export volume, and household consumption) and policy tools (which may be export tax rebate rates, loan interest rates, and fiscal subsidy amounts). In addition, edges represent causal relationships between nodes, and each edge has a causal type label, including "positive incentive", "negative inhibition" and "indirect adjustment". Here, positive incentive can be increasing export tax rebates leading to increased export volume, negative inhibition can be raising interest rates leading to curbing overheated investment, and indirect adjustment can be industrial guidance funds leading to changes in the output value of emerging industries, which further leads to changes in GDP. The construction of this knowledge graph employs the following steps: Causal relationships were extracted from historical policy texts, economic research reports, academic literature, and expert rule bases. Among them, entity extraction uses a pre-trained natural language processing model from existing technologies to perform named entity recognition, identifying the names of economic indicators and policy tools. Relation extraction employs a remote supervision method to identify the causal relationships and directions between entity pairs; The extracted results are stored in the graph database after being verified by experts. Furthermore, the knowledge graph supports periodic updates, so that newly released policy documents and research reports can be continuously processed and added to the graph.

[0053] Step S6: Substitute the current economic state, the set of combined intervention strategies, and the structural deviation vector into the differentiable structural causal model to generate a multi-dimensional combined intervention strategy and the time-series execution parameters of the multi-dimensional combined intervention strategy; In some embodiments, substituting the current economic state, the set of combined intervention strategies, and the structural deviation vector into a differentiable structural causal model to generate a multi-dimensional combined intervention strategy and the time-series execution parameters of the multi-dimensional combined intervention strategy includes: The current economic state and the structural deviation vector are concatenated to form a comprehensive economic state description vector, which is then used as the input to the policy network. The dimensions of this economic state description vector include the number of economic indicators and the deviation vector dimension. The economic state description vector is input into the policy network, and the policy network outputs candidate actions based on the current state. The candidate actions include at least the selection of intervention tools, the intensity parameter of the intervention tool, and the execution order type. The selection of intervention tools and the intensity parameters corresponding to the intervention tools in the candidate actions are mapped to the do operator intervention of the corresponding economic variables in the differentiable causal model. Based on the execution order type in the candidate action, the do operator intervention of the corresponding economic variable is carried out in the differentiable causal model to obtain the environmental output result. The environmental output result includes at least the predicted values ​​of each economic indicator in each future time step (which can be the next 4 to 8 quarters), and the corresponding immediate reward calculated based on each predicted value, that is, the comprehensive score of policy effect. The environmental output is combined with the evaluation of the value network to calculate the analytical gradient of the action value function with respect to the action. This analytical gradient includes the gradient of the immediate reward with respect to the action and the gradient of the environmental Jacobian matrix multiplied by the next state value. The analytical gradient is backpropagated to the policy network to update the network's parameters, making the network more inclined to output actions that yield higher cumulative rewards. During training, the weights of each objective are dynamically adjusted in the multi-objective reward function to obtain the optimal policy set. After multiple iterations, when the policy performance no longer improves after several consecutive iterations, i.e. when the cumulative reward change is less than a preset threshold, the training is terminated, and the optimal combination intervention policy generated by the current policy network and its time-series execution parameters are output.

[0054] In some embodiments, based on the execution order type in the candidate action, intervention is performed in the do operator of the corresponding economic variable in a differentiable causal model, including: If the execution order type is synchronous execution, then the do interventions of all selected tools are applied to the structural causal model at the same time step, and a forward propagation is performed to obtain the predicted values ​​of future economic indicators. If the execution order type is sequential, the structural causal model is expanded over time. The first tool's do-intervention is applied at the first time step, and the intermediate state is obtained through forward propagation. The second tool's do-intervention is applied at the next time step, and forward propagation continues. This process is repeated until all tools have been executed, and the predicted values ​​of future economic indicators are obtained. If the execution order type is staggered execution, each tool will be assigned a time offset parameter, which is a value between 0 and the total duration, to represent the delayed execution time of the tool relative to the start time. In the time unfolding of the structural causal model, interventions will be gradually applied at the corresponding time steps according to the time offset of each tool to obtain the predicted values ​​of future economic indicators.

[0055] In some embodiments, the policy network employs three parallel output heads that share an underlying economic state encoding network. This encoding network compresses the input vector into a fixed-dimensional feature representation (here, a 256-dimensional vector), and the three output heads generate three decision outputs based on this feature representation: The selection head is used to output the probability of each tool being selected. That is, for each tool in the candidate combination, it outputs a value between 0 and 1 to indicate the probability that the tool should be used. The strength head is used to output the specific strength value of each tool, that is, to output the mean and standard deviation of the implementation strength of each selected tool. The mean determines the specific strength value, and the standard deviation controls the exploration degree. That is, the output strength value is mapped to the actual strength range of each tool through mathematical transformation. The sequence header is used to output the sorting score matrix between the tools; this matrix is ​​a... The matrix is ​​a square matrix, where the element in the i-th row and j-th column represents the probability score of tool i being ranked before tool j. The discrete choices are made continuous by the Gumbel-Softmax relaxation method, which transforms the discrete order selection into a differentiable problem. The sorting score matrix is ​​then transformed into an effective execution order arrangement by combining the Hungarian algorithm. The true value of the gradient of the action value function is calculated based on the differentiable causal model, which reduces variance and improves sample efficiency. Specifically, this method can converge within tens to hundreds of rounds. The reward function can be a multi-objective weighted form, and the weight coefficients can be fixed or dynamically adjusted during training to achieve Pareto optimal search.

[0056] In some embodiments, the formula for calculating the policy gradient is:

[0057] in, For policy networks, based on the current economic state Output intervention actions ; For policy network parameters The gradient; Action value function Action The gradient; It represents the expectation of a state and an action.

[0058] Gradient expansion of the action value function:

[0059] in, For instant reward functions; For a differentiable causal model, the current state is the input. and actions Output the next state ; The Jacobian matrix of the environment model; This is the discount factor (usually taken as 0.95~0.99). The value gradient for the next state.

[0060] In some embodiments, the reward function is a multi-objective weighted function that integrates economic growth, employment stability, price stability, external equilibrium, and policy feasibility constraints.

[0061] Step S7: Perform counterfactual simulation verification on the multi-dimensional combined intervention strategy, and output the intervention effect prediction and expected causal chain.

[0062] In some embodiments, the expected causal chain is visualized as a directed graph, with an estimated causal effect and confidence interval for each edge; simultaneously, a counterfactual comparison table is output to compare the differences in the predicted trajectories of key economic indicators under two scenarios: no intervention and the adoption of the adaptive intervention strategy.

[0063] In some embodiments, the specific steps of the counterfactual simulation verification include: Starting from the current economic situation, set up at least two counterfactual scenarios: Scenario 1 is set as a non-intervention scenario, that is, no additional policy intervention is taken, and the system is left to evolve on its own. Scenario 2 is the multi-dimensional combined intervention strategy.

[0064] In addition, comparative scenarios such as "adopting alternative strategies" and "adopting strategies of different intensities" can be set up. For each scenario, forward propagation prediction is performed using a differentiable structural causal model to simulate the evolution trajectory of each key economic indicator over the next N time steps (which can be 4 to 8 quarters). Since the structural equations in the differentiable structural causal model may contain randomness, such as the distribution of latent variables, multiple samplings are performed using the Monte Carlo method, which is 1000 times in this case, to obtain the distribution of each predicted indicator.

[0065] The output may include: The complete causal path from the intervention node, through the mediating variable, to the target variable is presented in the form of a directed graph. Each directed edge is labeled with the estimated causal effect and the 95% confidence interval to make the expected causal chain visible, thereby enabling decision-makers to intuitively understand the effect and degree of the strategy. By comparing the key economic indicators, namely GDP growth rate, employment rate, CPI, and trade balance, at each forecast time step in the two scenarios above in tabular form, a counterfactual comparison table is generated, which simultaneously indicates whether the differences between the two scenarios are statistically significant. For this multidimensional combined intervention strategy, the median predicted value, 25%-75% quantile range and 90% confidence interval of each target indicator during the prediction period are output, along with the source analysis of prediction uncertainty, such as the uncertainty of implicit variables and / or the uncertainty from structural equation estimation error, to reflect the prediction range of intervention effect and uncertainty measurement. The above outputs provide decision-makers with sufficient evidence to assess whether to adopt the strategy, whether to adjust the strategy strength, or to choose an alternative strategy.

[0066] Furthermore, the present invention also provides a fluctuation attribution and adaptive intervention strategy generation system based on SCM.

[0067] See appendix Figure 4 , Figure 4 This is a schematic diagram of the main structure of a fluctuation attribution and adaptive intervention strategy generation system based on SCM according to an embodiment of the present invention. Figure 4 As shown, an embodiment of the present invention provides a SCM-based fluctuation attribution and adaptive intervention strategy generation system for implementing any of the above-described methods, comprising the following steps: The data acquisition module is used to acquire multi-source macroeconomic time-series data and preprocess the multi-source macroeconomic time-series data to obtain multi-source macroeconomic time-series data of the target region. The differentiable causal discovery module is used to obtain a differentiable structural causal model based on multi-source macroeconomic time-series data of the target region. The detection module is used to monitor the fluctuations of the target economic indicator in real time, and based on the fluctuations of the target indicator, detect whether the target economic indicator deviates from the preset trajectory. The causal attribution module is used to perform multivariate counterfactual perturbation analysis based on the differentiable structural causal model when the target economic indicator deviates from the expected trajectory, identify the core driving factors and generate a structural deviation vector. The combined intervention strategy generation module is used to generate a set of combined intervention strategies based on the core driving factors, the structural deviation vector, and the pre-constructed intervention anchor knowledge graph. The intervention anchor knowledge graph module is used to store and manage the causal relationship knowledge between economic fluctuation types, driving factor categories and policy tools, so as to facilitate retrieval and reasoning based on the semantics of candidate intervention anchors; The combined intervention optimization module is used to simultaneously solve the joint optimal solution of three dimensions—intervention tool selection, intervention intensity parameter, and execution order type—based on the current economic state, the combined intervention strategy set, and the structural deviation vector, through the Jacobian matrix of the differentiable structural causal model, and generate a multi-dimensional combined intervention strategy and the time-series execution parameters of the multi-dimensional combined intervention strategy. The counterfactual simulation and interpretable output module is used to perform multi-scenario counterfactual simulations on the multi-dimensional combined intervention strategy, and output the intervention effect prediction range, causal chain visualization and decision suggestions.

[0068] In some embodiments, the acquisition module includes: The data acquisition unit is used to collect multi-source macroeconomic time-series raw data for the target region. The natural language processing unit is used to quantify policy texts such as government work reports, central bank monetary policy implementation reports, and fiscal budget reports to generate policy text quantitative indices. The policy text quantitative indices include at least the following dimensions: policy direction coding category (which can be easing / tightening / neutral), policy strength score, and policy implementation time window. The data preprocessing unit is used to fill in missing values ​​using linear interpolation, identify and process outliers using the 3σ principle, and perform Z-score standardization on each indicator to eliminate the impact of dimensional differences on subsequent causal analysis.

[0069] In some embodiments, the system further includes a robust screening module, the robust screening module comprising: The Bayesian posterior computation unit is used to perform weighted resampling on the original dataset in the differentiable structural causal model to obtain several weighted sampled datasets, and to re-estimate the causal graph structure and structural equation parameters on each weighted sampled dataset to generate several posterior causal graph samples and corresponding structural equation parameter samples. The average causal unit is used to construct a set of candidate core driving factors and a posterior differentiable causal model based on each posterior causal graph sample, and to determine the average causal effect of each candidate core driving factor on the target variable and the posterior distribution of each candidate core driving factor based on the set of candidate core driving factors and the posterior distribution of each candidate core driving factor. The confidence index calculation unit is used to determine the causal effect confidence index of each candidate core driver factor based on the average causal effect of each candidate core driver factor on the target variable and the posterior distribution of each candidate core driver factor. The driving factor screening unit is used to screen the causal effect confidence index of each candidate core driving factor to obtain robust core driving factors and the factors to be observed.

[0070] In some embodiments, the combined intervention strategy generation module includes: The knowledge graph matching unit is used to determine the corresponding economic variable node of the robust core driving factor in the knowledge graph based on the robust core driving factor, and to perform a reverse breadth-first search in the knowledge graph based on each economic variable node as the target node to determine the policy tool node of each robust core driving factor and obtain all the influencing parameters. An intervention action generation unit is used to generate several candidate intervention actions and the corresponding time-series execution parameters for each candidate intervention action based on all influencing parameters and the structural deviation vector. The combined strategy generation unit is used to determine the priority of each candidate intervention action based on several candidate intervention actions and the timing execution parameters corresponding to each candidate intervention action, and to construct a combined intervention strategy set according to the priority of each candidate intervention action.

[0071] In some embodiments, the differentiable causality discovery module includes: The dynamic structural causal model generation unit is used to construct dynamic structural causal models based on multi-source macroeconomic time-series data and time-series causal discovery algorithms for latent variables. The training unit is used to train the dynamic structural causal model based on multi-source macroeconomic time series data to obtain the trained dynamic structural causal model, which serves as a differentiable structural causal model.

[0072] In some embodiments, the differentiable causality discovery module includes: Lag analysis unit: used to determine the time lag relationship between various economic variables; Differentiable graph structure learning unit: The NOTEARS or DAG-GNN algorithm is used to learn the directed acyclic graph structure through continuous optimization. Latent variable introduction unit: Used to add latent variable nodes representing unobserved factors in a cause-effect graph; Neural structural equation unit: parameterizes each conditional probability distribution into a differentiable neural network.

[0073] In some embodiments, the causal attribution module includes: The counterfactual perturbation sampling unit is used to replace the parent node with counterfactual values ​​at several different quantiles; The contribution determination unit is used to run the do operator intervention based on the replaced parent node and obtain the structural contribution of the parent node; The core driving factor determination unit is used to identify the core driving factors based on each structural contribution. The causal baseline determination unit is used to determine the causal baseline value of the core driving factor by taking the counterfactual expectation value. The deviation vector output unit is used to output a structural deviation vector based on the current value of each core driving factor and its Ginza benchmark value.

[0074] In some embodiments, the intervention anchor knowledge graph module includes: The entity extraction unit is used to extract economic variables and policy tool entities from policy documents and economic research reports; The causal relationship extraction unit is used to extract causal relationships and type labels between entities based on the natural language processing model. The graph storage unit is used to store knowledge graphs with causal edge types in the form of a graph database; The reasoning retrieval unit is used to retrieve matching candidate intervention anchors and their expected causal paths based on the input core driving factors and fluctuation types.

[0075] In some embodiments, the combined intervention optimization module includes: The economic state description vector generation unit is used to generate a comprehensive economic state description vector based on the current economic state and the structural deviation vector. The policy network unit is used to construct a shared underlying economic state encoding network with three parallel output heads. The economic state description vector is input into the policy network to output candidate actions. The candidate action intervention unit is used to intervene in the corresponding economic variable in the differentiable causal model according to the execution order type of the candidate actions, obtain the environmental output result, and combine the environmental output result with the evaluation of the value network to calculate the analytical gradient of the action value function with respect to the action. The iterative optimization unit is used to backpropagate the analytical gradient to the policy network, update the parameters of the policy network, and perform multiple iterations until training is terminated. Then, it outputs the optimal combination intervention policy and its time-series execution parameters generated by the current policy network.

[0076] In some embodiments, the counterfactual simulation and interpretable output module includes: The scenario building unit is used to construct multiple counterfactual scenarios to facilitate counterfactual simulation and verification. The simulation verification unit is used to perform counterfactual simulation verification according to different scenarios, and output the corresponding intervention effect prediction and expected causal chain.

[0077] In practical deployment, the system of this invention can be implemented through a microservice architecture or an integrated software platform: Specifically, the data acquisition module runs as a scheduled task, automatically pulling the latest data and performing preprocessing daily or quarterly to update the local database; the differentiable causal discovery module runs during system initialization or triggers retraining after receiving a sufficient amount of new data, outputting a differentiable causal model that is persistently stored and available for use by other modules; the causal attribution module runs at a high frequency, possibly daily, monitoring the fluctuations of target indicators in real time, and immediately performs attribution analysis and pushes the results to the intervention anchor knowledge graph module once a threshold is triggered; the intervention anchor knowledge graph module is deployed independently as a static knowledge base and supports online updates; the combined intervention optimization module is started on demand after the attribution results are generated and completes strategy optimization within minutes; the counterfactual simulation and interpretable output module is executed last to generate a decision report.

[0078] The aforementioned SCM-based volatility attribution and adaptive intervention strategy generation system is used for implementation. Figure 1 The embodiments of the SCM-based fluctuation attribution and adaptive intervention strategy generation method shown are similar in technical principle, the technical problems solved, and the technical effects produced. Those skilled in the art can clearly understand that, for the sake of convenience and brevity, the specific working process and related descriptions of the SCM-based fluctuation attribution and adaptive intervention strategy generation system can be found in the embodiments of the SCM-based fluctuation attribution and adaptive intervention strategy generation method, and will not be repeated here.

[0079] It should be noted that although the steps in the above embodiments are described in a specific order, those skilled in the art will understand that in order to achieve the effects of the present invention, different steps do not necessarily have to be executed in such an order. They can be executed simultaneously (in parallel) or in other orders, and these variations are all within the scope of protection of the present invention.

[0080] Those skilled in the art will understand that the implementation of all or part of the operational flow of the SCM-based fluctuation attribution and adaptive intervention strategy generation system of the present invention can also be accomplished by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium. When executed by a processor, the computer program can implement the steps of the above-described embodiment of the operational flow of the SCM-based fluctuation attribution and adaptive intervention strategy generation system. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or some intermediate form. The computer-readable storage medium can include any entity or device capable of carrying the computer program code, a medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory, a random access memory, an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc. It should be noted that the content included in the computer-readable storage medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the computer-readable storage medium does not include electrical carrier signals and telecommunication signals.

[0081] Furthermore, the SCM-based fluctuation attribution and adaptive intervention strategy generation system of the present invention also includes a computer-readable storage medium. In one embodiment of the present invention, the computer-readable storage medium can be configured to store a program executing the above-described SCM-based fluctuation attribution and adaptive intervention strategy generation system embodiment. This program can be loaded and run by a processor to implement the operation flow of the above-described SCM-based fluctuation attribution and adaptive intervention strategy generation system. For ease of explanation, only the parts related to the embodiments of the present invention are shown; for specific technical details not disclosed, please refer to the SCM-based fluctuation attribution and adaptive intervention strategy generation system section of the embodiments of the present invention. The computer-readable storage medium can be a storage device comprising various electronic devices. Optionally, in the embodiments of the present invention, the computer-readable storage medium is a non-transitory computer-readable storage medium.

[0082] Furthermore, it should be understood that since the various modules are only provided to illustrate the functional units of the device of the present invention, the physical devices corresponding to these modules may be the processor itself, or a part of the processor's software, hardware, or a combination of software and hardware. Therefore, the number of modules shown in the figures is merely illustrative.

[0083] Those skilled in the art will understand that the various modules in the device can be adaptively split or combined. Such splitting or combining of specific modules will not cause the technical solution to deviate from the principles of the present invention; therefore, the technical solutions after splitting or combining will fall within the protection scope of the present invention.

[0084] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A method for generating volatility attribution and adaptive intervention strategies based on SCM, characterized in that, Includes the following steps: Acquire multi-source macroeconomic time-series data for the target region; Based on the aforementioned multi-source macroeconomic time-series data, a differentiable causal model is constructed. Real-time monitoring to see if target economic indicators deviate from the preset trajectory: If the target economic indicator deviates from the expected trajectory, a multivariate counterfactual disturbance analysis is performed based on the differentiable structural causal model to identify the core driving factors and generate a structural deviation vector. Based on the core driving factors, the structural deviation vector, and the pre-built intervention anchor knowledge graph, a set of combined intervention strategies is generated. The set of combined intervention strategies includes at least two candidate intervention actions of different action types and the time-series execution parameters corresponding to each candidate intervention action. Substituting the current economic state, the set of combined intervention strategies, and the structural deviation vector into a differentiable structural causal model, a multi-dimensional combined intervention strategy and its time-series execution parameters are generated. The multi-dimensional combined intervention strategy is verified by counterfactual simulation, and the predicted intervention effect and expected causal chain are output.

2. The method for generating fluctuation attribution and adaptive intervention strategies based on SCM according to claim 1, characterized in that, The construction of a differentiable structural causal model based on the multi-source macroeconomic time-series data includes: Based on the aforementioned multi-source macroeconomic time-series data, a dynamic structural causal model is constructed, wherein the dynamic structural causal model includes observable variables and latent variable nodes; The dynamic structural causal model is trained to obtain a differentiable structural causal model.

3. The method for generating fluctuation attribution and adaptive intervention strategies based on SCM according to claim 2, characterized in that, The multi-source macroeconomic time-series data includes at least: regional GDP, industrial output, consumption indicators, investment indicators, import and export indicators, money supply, price index, and quantitative index based on policy documents.

4. The method for generating fluctuation attribution and adaptive intervention strategies based on SCM according to claim 1, characterized in that, The step of performing multivariate counterfactual perturbation analysis based on the differentiable structural causal model to identify core driving factors and generate a structural bias vector includes: Using the target economic indicator as the target variable, obtain the set of parent nodes of the target variable in the differentiable causal model. For each parent node in the set of parent nodes, fix the other parent nodes as the current true observation value, and replace the parent node with several counterfactual values ​​under different quantiles; Based on the differentiable structural causal model, the do operator intervention is performed to identify the core driving factors and generate a structural deviation vector.

5. The method for generating fluctuation attribution and adaptive intervention strategies based on SCM according to claim 4, characterized in that, The step of performing do operator intervention based on the differentiable structural causal model to identify core driving factors and generate a structural deviation vector includes: Based on the replaced parent node, the do operator intervention is performed based on the differentiable causal model to obtain the expected value of the target variable under each counterfactual value intervention; The average of the absolute differences between the expected value of the target variable under all quantiles and the expected value of the target variable when all parent nodes take the current true value is taken as the structural contribution of the parent node. The top few factors with the largest structural contributions are identified as the core driving factors; For each core driving factor, the mean of the historical stationary period of the parent node of the core driving factor is taken as the stationary state value. Based on the differentiable causal model, the counterfactual expected value of the core driving factor when the parent node takes the steady state value is determined, and the counterfactual expected value is used as the causal benchmark value of the core driving factor. The difference between the current value of each core driving factor and the causal baseline value of the core driving factor is divided by the standard deviation of the core driving factor during historical stationary periods to obtain the standardized deviation of each core driving factor. All standardized deviations are combined as the structural deviation vector.

6. The method for generating fluctuation attribution and adaptive intervention strategies based on SCM according to claim 1, characterized in that, Before generating the combined intervention strategy set based on the core driving factors, the structural deviation vector, and the pre-constructed intervention anchor knowledge graph, the method further includes: Bayesian posterior calculations are performed on differentiable structural causal models to generate several posterior causal graph samples and corresponding structural equation parameter samples. Based on each posterior causal graph sample, determine the average causal effect of each candidate core driver on the target variable and the posterior distribution of each candidate core driver. Based on the average causal effect of each candidate core driver factor on the target variable and the posterior distribution of each candidate core driver factor, the causal effect confidence index of each candidate core driver factor is determined. The causal effect confidence index of each candidate core driver factor is used to screen for robust core drivers factors and factors to be observed.

7. The method for generating fluctuation attribution and adaptive intervention strategies based on SCM according to claim 6, characterized in that, The process of generating a set of combined intervention strategies based on the core driving factors, the structural deviation vector, and the pre-constructed intervention anchor knowledge graph includes: Based on the robust core driving factors, the pre-constructed intervention anchor knowledge graph is matched to obtain all influencing parameters; Based on all influencing parameters and the structural deviation vector, several candidate intervention actions and the corresponding time-series execution parameters for each candidate intervention action are generated. Based on several candidate intervention actions and the timing execution parameters corresponding to each candidate intervention action, a set of combined intervention strategies is constructed.

8. A fluctuation attribution and adaptive intervention strategy generation system based on SCM, used to implement the steps of the method according to any one of claims 1-7, characterized in that, The system includes: The data acquisition module is used to acquire multi-source macroeconomic time-series data and preprocess the multi-source macroeconomic time-series data to obtain multi-source macroeconomic time-series data of the target region. The differentiable causal discovery module is used to obtain a differentiable structural causal model based on multi-source macroeconomic time-series data of the target region. The detection module is used to monitor the fluctuations of the target economic indicator in real time, and based on the fluctuations of the target indicator, detect whether the target economic indicator deviates from the preset trajectory. The causal attribution module is used to perform multivariate counterfactual perturbation analysis based on the differentiable structural causal model when the target economic indicator deviates from the expected trajectory, identify the core driving factors and generate a structural deviation vector. The combined intervention strategy generation module is used to generate a set of combined intervention strategies based on the core driving factors, the structural deviation vector, and the pre-constructed intervention anchor knowledge graph. The intervention anchor knowledge graph module is used to store and manage the causal relationship knowledge between economic fluctuation types, driving factor categories and policy tools, so as to facilitate retrieval and reasoning based on the semantics of candidate intervention anchors; The combined intervention optimization module is used to simultaneously solve the joint optimal solution of three dimensions—intervention tool selection, intervention intensity, and execution order—based on the current economic state, the set of combined intervention strategies, and the structural deviation vector, through the Jacobian matrix of the differentiable structural causal model, and generate a multi-dimensional combined intervention strategy and the time-series execution parameters of the multi-dimensional combined intervention strategy. The counterfactual simulation and interpretable output module is used to perform multi-scenario counterfactual simulations on the multi-dimensional combined intervention strategy, and output the intervention effect prediction range, causal chain visualization and decision suggestions.

9. The SCM-based fluctuation attribution and adaptive intervention strategy generation system according to claim 8, characterized in that, The system also includes a robust screening module, which comprises: The Bayesian posterior computation unit is used to perform weighted resampling on the original dataset in the differentiable structural causal model to obtain several weighted sampled datasets, and to re-estimate the causal graph structure and structural equation parameters on each weighted sampled dataset to generate several posterior causal graph samples and corresponding structural equation parameter samples. The average causal unit is used to construct a set of candidate core driving factors and a posterior differentiable causal model based on each posterior causal graph sample, and to determine the average causal effect of each candidate core driving factor on the target variable and the posterior distribution of each candidate core driving factor based on the set of candidate core driving factors and the posterior distribution of each candidate core driving factor. The confidence index calculation unit is used to determine the causal effect confidence index of each candidate core driver factor based on the average causal effect of each candidate core driver factor on the target variable and the posterior distribution of each candidate core driver factor. The driving factor screening unit is used to screen the causal effect confidence index of each candidate core driving factor to obtain robust core driving factors and the factors to be observed.

10. The SCM-based fluctuation attribution and adaptive intervention strategy generation system according to claim 9, characterized in that, The combined intervention strategy generation module includes: The knowledge graph matching unit is used to determine the corresponding economic variable node of the robust core driving factor in the knowledge graph based on the robust core driving factor, and to perform a reverse breadth-first search in the knowledge graph based on each economic variable node as the target node to determine the policy tool node of each robust core driving factor and obtain all the influencing parameters. An intervention action generation unit is used to generate several candidate intervention actions and the corresponding time-series execution parameters for each candidate intervention action based on all influencing parameters and the structural deviation vector. The combined strategy generation unit is used to determine the priority of each candidate intervention action based on several candidate intervention actions and the timing execution parameters corresponding to each candidate intervention action, and to construct a combined intervention strategy set according to the priority of each candidate intervention action.

Images