Agricultural product price data hedging method based on electric data processing

Abstract

The invention provides an agricultural product price data hedging method based on electric data processing, and the method comprises the steps: collecting and fusing the multi-source data of agricultural product prices, meteorology, policies, supply chains, macroeconomy and the like, and constructing an evolutionary dynamic causal graph in combination with time sequence synchronization, standardized processing and causal relationship mining; a dual-channel graph neural network is introduced to realize strategy generation and interpretability path synchronous reasoning, a reinforcement learning algorithm is utilized to output optimal hedging operation, decision logic is converted into natural language interpretation, consistency and traceability of strategies and interpretation are ensured, intelligence, adaptability and transparency of hedging decision are improved, and the method is suitable for popularization and application. And the risk management challenge caused by market fluctuation can be dealt with.

Description

Technical Field

[0001] This invention relates to the field of market data analysis and intelligent decision support technology, and in particular to a method for hedging agricultural product price data using electronic data processing. Background Technology

[0002] Existing agricultural product price risk management and hedging strategy recommendation technologies are primarily price data-driven, employing machine learning, time series analysis, and expert rule models to model and predict market fluctuations. Mainstream methods often extract features from historical prices, climate, policy information, and supply chain indicators, then use deep neural networks, regression models, or ensemble learning methods to generate hedging solutions, further enhanced by knowledge graphs, association rules, or causal networks to improve interpretability. Some technological systems attempt to achieve knowledge-driven intelligent decision-making. In existing technologies, strategy recommendation and strategy explanation modules are typically implemented separately. The recommendation process focuses on outputting the optimal hedging action sequence, while the explanation module independently provides a source description of the solution's outcome based on a knowledge graph or causal relationships. This structure results in a logical disconnect between the explanation content and the strategy recommendation, failing to guarantee a strict causal alignment between the explanation text and the actual recommended actions, thus affecting the overall credibility and traceability of the system. Typical examples include agricultural product price forecasting models based on time series analysis, which use ARIMA, LSTM, or VAR models to capture market trends and are suitable for price trend prediction and routine risk management scenarios. Existing technologies also propose explanation enhancement techniques based on causal inference and knowledge graphs, which can assist users in understanding recommendation results and their decision-making basis to some extent. However, these often involve retrospective explanations of existing decisions after the recommendation has been made, failing to achieve real-time synchronization between strategy actions and explanation paths. Summary of the Invention

[0003] In order to solve the above-mentioned technical problems, the present invention provides a method for hedging agricultural product price data through electronic data processing.

[0004] The technical solution of this invention is implemented as follows: a method for hedging agricultural product price data through electronic data processing, comprising: S1: Obtain historical price series of agricultural products, climate data, policy announcements, supply chain logistics information and macroeconomic indicators as multi-source raw input data to construct a set of external influencing factors for hedging strategy generation; S2: Perform time alignment and standardization preprocessing on the multi-source raw input data. Based on Granger causality test and time-series nonlinear dependency analysis, identify the dynamic driving relationship between variables and generate an initial dynamic causal graph structure, where nodes represent price fluctuation-related factors and edges represent statistically verified causal directions and strengths. S3: Using the initial dynamic causal graph structure as the basic reasoning framework, an online learning mechanism is introduced to adjust the weight parameters and conditional probability distribution of the causal edges according to the real-time updated market feedback data, generating an evolutionary dynamic causal graph model with environmental adaptability, which is used to characterize the price influence mechanism that changes over time. S4: Based on the evolutionary dynamic causal graph model, a dual-channel graph neural network architecture is constructed, in which the policy generation channel and the explanation generation channel share the same graph structure representation, and the synchronous attention gating mechanism ensures that the two channels maintain state consistency in each reasoning step, forming a joint decision-explanation reasoning space; S5: In the joint decision-explanation reasoning space, the reinforcement learning algorithm is used to perform the optimal hedging action search with the objective function of minimizing the expected price risk, generating a strategy operation sequence, and recording the set of core causal paths activated in the process. S6: Input the core causal path set into the semantic mapping module, convert it into a logical chain description readable by natural language based on the predefined causal semantic rule library, generate interpretable text fragments that strictly correspond to the policy actions, and constitute the policy explanation output; S7: Determine whether the logical consistency threshold requirement is met between the strategy operation sequence and the corresponding interpretable text fragment. If not, trigger the causal subgraph recalibration mechanism to re-retrieve or optimize the relevant causal path based on the latest market state until a consensus is reached. S8: Output the final hedging solution, which includes hedging operation instructions that have been verified for consistency and a complete chain of decision-making basis, realizing a closed-loop collaborative output of strategy recommendation and interpretation generation under a unified model.

[0005] The present invention provides a method for hedging agricultural product price data through electronic data processing, which has the following beneficial effects: (1) This invention achieves synchronous coordination between policy decision-making and interpretation output in terms of temporal evolution and semantic dependence by constructing a dual-channel graph neural network architecture with a shared dynamic causal graph structure. The policy generation channel and interpretation generation channel share the same underlying causal graph, and an attention gating mechanism is used to align states at each inference step, ensuring that each operation instruction is generated based on the most relevant causal propagation path at the current moment. Simultaneously, the interpretation channel extracts this path and converts it into a natural language description, fundamentally avoiding the technical defect of policy and interpretation being disconnected. A consistency verification module is introduced to evaluate the logical matching degree of each output action-interpretation pair. If a causal chain break or weight deviation is detected, a recalibration mechanism is triggered to dynamically reconstruct the subgraph, thereby significantly improving the credibility and traceability of the system output. (2) This invention integrates a time-series causal discovery algorithm with an online learning mechanism to construct a causal reasoning framework with dynamic adaptability, effectively improving the system's responsiveness and predictive robustness to changes in the external environment. First, based on Granger causality tests and nonlinear dependency analysis, the key driving factors of price fluctuations and their directions of action are identified, forming an initial causal graph with clear economic meaning. Then, combined with real-time collected data such as price feedback, weather updates, and policy releases, an online parameter estimation method is used to continuously adjust the strength weights and conditional probability distributions of the causal edges, enabling the model to reflect structural changes brought about by sudden shocks (such as extreme weather or changes in trade policy) in a timely manner. This mechanism not only enhances the stability of the system in non-stationary market environments but also ensures that the generated hedging strategies are always based on the latest and most relevant causal evidence, significantly improving the foresight and accuracy of decision-making. Attached Figure Description

[0006] Figure 1 A flowchart of an agricultural product price data hedging method for electrical data processing according to the present invention; Figure 2 This is a sub-flowchart of an agricultural product price data hedging method for electronic data processing according to the present invention; Figure 3 This is another sub-flowchart of an agricultural product price data hedging method for electronic data processing according to the present invention. Detailed Implementation

[0007] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0008] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0009] like Figure 1 As shown, this invention provides a method for hedging agricultural product price data using electronic data processing, specifically including: S1: Obtain historical price series of agricultural products, climate data, policy announcements, supply chain logistics information and macroeconomic indicators as multi-source raw input data to construct a set of external influencing factors for hedging strategy generation; S2: Perform time alignment and standardization preprocessing on the multi-source raw input data. Based on Granger causality test and time-series nonlinear dependency analysis, identify the dynamic driving relationship between variables and generate an initial dynamic causal graph structure, where nodes represent price fluctuation-related factors and edges represent statistically verified causal directions and strengths. S3: Using the initial dynamic causal graph structure as the basic reasoning framework, an online learning mechanism is introduced to adjust the weight parameters and conditional probability distribution of the causal edges according to the real-time updated market feedback data, generating an evolutionary dynamic causal graph model with environmental adaptability, which is used to characterize the price influence mechanism that changes over time. S4: Based on the evolutionary dynamic causal graph model, a dual-channel graph neural network architecture is constructed, in which the policy generation channel and the explanation generation channel share the same graph structure representation, and the synchronous attention gating mechanism ensures that the two channels maintain state consistency in each reasoning step, forming a joint decision-explanation reasoning space; S5: In the joint decision-explanation reasoning space, the reinforcement learning algorithm is used to perform the optimal hedging action search with the objective function of minimizing the expected price risk, generating a strategy operation sequence, and recording the set of core causal paths activated in the process. S6: Input the core causal path set into the semantic mapping module, convert it into a logical chain description readable by natural language based on the predefined causal semantic rule library, generate interpretable text fragments that strictly correspond to the policy actions, and constitute the policy explanation output; S7: Determine whether the logical consistency threshold requirement is met between the strategy operation sequence and the corresponding interpretable text fragment. If not, trigger the causal subgraph recalibration mechanism to re-retrieve or optimize the relevant causal path based on the latest market state until a consensus is reached. S8: Output the final hedging solution, which includes hedging operation instructions that have been verified for consistency and a complete chain of decision-making basis, realizing a closed-loop collaborative output of strategy recommendation and interpretation generation under a unified model.

[0010] Step S1 involves acquiring historical price series of agricultural products, climate data, policy announcements, supply chain logistics information, and macroeconomic indicators as multi-source raw input data to construct a set of external influencing factors for hedging strategy generation. Specifically, this includes: S1.1: Obtain historical price series of agricultural products and daily meteorological observation data from national agricultural data centers, futures exchanges and meteorological bureaus, and perform cross-source time series synchronization processing based on timestamp alignment rules to generate a time-aligned dataset with a uniform sampling frequency, which serves as the basic input for subsequent causal analysis; To acquire historical price series of agricultural products from the national agricultural data center and daily market data from futures exchanges, a cross-source data interface call mechanism (parameters: API address, authentication key, time granularity) is adopted to achieve automated batch collection of high-precision price time series. Furthermore, through the meteorological bureau's observation data interface (parameters: observation station number, sampling interval, meteorological parameter type), daily meteorological observation data, including indicators such as temperature, precipitation, humidity, and wind speed, are obtained to form a raw set of meteorological time series data. Furthermore, by using timestamp alignment rules (parameters: unified time zone, time format ISO8601, and fixed sampling frequency Δt), and employing time index-based resampling and linear interpolation algorithms, synchronization processing is performed on the historical price series of agricultural products and the meteorological observation time series to eliminate cross-source sampling frequency differences. Furthermore, a sliding window segmentation method (parameters: window length W, step size S) is adopted to divide the time series into fixed-length analysis segments based on a unified sampling frequency, and the mean and variance within the window are calculated to capture short-term fluctuation characteristics. Furthermore, an outlier detection algorithm (parameters: 3σ principle threshold, MAD median absolute deviation threshold) is used to screen the time series segments for quality, remove data points that significantly deviate from the normal range, and generate a time-aligned dataset without outliers; Through time series synchronization processing and quality screening chain algorithm, the original price data and meteorological observation data collected from multiple sources are transformed into a multi-source time-aligned dataset with consistent sampling frequency and balanced quality, so as to realize a unified time reference input for subsequent causal analysis; For example, in a corn futures hedging application scenario, the API call time granularity is set to 1 day, Δt is 24 hours, and corn spot and futures trading prices are collected from an agricultural data center, with the price unit being yuan / ton. Meteorological observation data includes daily average temperature, cumulative rainfall, and average wind speed from three observation stations in the main producing area. During time alignment, the data is uniformly converted to Beijing time and resampled at a frequency of Δt=24 hours. Missing data is interpolated using linear interpolation: for example, between two known observations... Between degrees Celsius, the interpolated value for the day is Temperature in Celsius. With a window length W of 30 days and a step size S of 5 days, calculate the mean and standard deviation of prices for each window. The standard deviation formula is: ,in For the number of data points in the window, Price per point The window mean is used. Outlier detection employs the 3σ principle; any single point whose price exceeds the window mean ± 3σ is removed. The resulting time-aligned dataset significantly improves continuity and completeness, providing stable input for subsequent causal relationship modeling. S1.2: Collect supply chain logistics information from publicly released government agricultural policy announcements and customs import and export supervision records. Use named entity recognition and event extraction algorithms in natural language processing to extract key policy implementation nodes and logistics disruption events, and generate structured event tag sequences to characterize institutional and circulation exogenous shocks. Based on timestamp-aligned meteorological observation data and agricultural product price series as external references, a combination of web crawling and data interface calls is used to retrieve agricultural policy announcements and import and export logistics records from the official websites of government agricultural authorities and customs supervision systems, thereby achieving batch collection of policy-related and circulation-related external event data. Furthermore, through text preprocessing methods (parameters: UTF-8 encoding conversion, stop word filtering, sentence segmentation and word segmentation), the structured segmentation of announcement and record texts is achieved, forming a high-quality corpus sequence for subsequent algorithm input; Furthermore, a named entity recognition algorithm (parameters: BiLSTM-CRF model, pre-trained word vector dimension 300) is adopted to automatically identify entities in the text such as agricultural policy issuing agencies, policy implementation time, covered crop types, customs checkpoint names, and transportation vehicle types, and generate a set of entity labels with location indexes; Furthermore, by utilizing an event extraction algorithm (parameter: event trigger word detection and role filling rule base based on dependency syntax), the trigger word location and participating role parsing of key policy implementation nodes and supply chain disruption events are realized, resulting in an event triple set <entity1, event type, entity2>, which serves as the basic unit of the structured event chain; Furthermore, by using a time series embedding method (parameters: sliding window size of 7 days, window step size of 1 day), event triples are mapped to a unified time series coordinate system, achieving cross-domain time alignment of institutional exogenous shocks and liquidity exogenous shocks; By using an event encoding and labeling mechanism (parameters: event type encoding length 8 bits, time sequence index encoding length 4 bits), the result of the previous step is transformed into a machine-readable structured event labeling sequence, realizing the quantitative representation of institutional and circulation exogenous shocks, and providing a highly consistent input format for subsequent multi-source feature fusion; For example, in the agricultural data monitoring system, the API for government agricultural department announcements was called to obtain policy release texts from the past three years, totaling 215 texts, with an average length of 450 words per text. In the text preprocessing stage, a sentence segmentation algorithm was used to break long sentences into clauses with an average length of 15 words, and stop word filtering reduced the proportion of noisy words to 2%. In the named entity recognition process, the model parameters were set to a word vector dimension of 300, a BiLSTM hidden layer unit number of 256, and a CRF transition matrix dimension equal to the label set size of 9, identifying core entities such as policy implementation time, crop type, and issuing agency, totaling 1520 identified entities. In the event extraction stage, the rule base contained 36 types of agricultural policy events and 28 types of logistics anomaly events. Dependency parsing used the Arc-eager algorithm, successfully extracting 324 key policy implementation nodes and 198 logistics interruption events, forming a set of event triples. In the time embedding stage, a sliding window method is used to map events to a unified time-series coordinate system. The window size is set to 7 days, and the step size is 1 day, which aligns the event markers with the price series and meteorological data in terms of time steps, ultimately resulting in a structured event marker sequence with a length of 1095 steps. This sequence serves as the quantitative input for institutional and liquidity exogenous shocks and enters the subsequent multi-source feature fusion stage (S1.4). In the fusion verification stage, it demonstrates a significant effect on improving the node correlation in the causal graph construction stage. S1.3: Obtain macroeconomic indicators such as quarterly GDP growth rate, inflation rate and exchange rate fluctuations from the interface between the World Bank and the National Bureau of Statistics. Perform Z-score standardization on the above indicators to eliminate differences in dimensions and generate a normalized economic environment state vector as a quantitative representation of systemic risk factors. S1.4: The time-aligned meteorological observation data, structured event label sequence and normalized economic environment state vector are fused at the feature level. A high-dimensional joint feature matrix is ​​constructed based on the multi-source heterogeneous data splicing protocol to generate the original input tensor containing price-driven multi-dimensional factors. S1.5: Perform multiple imputation of missing values ​​and robust shrinking of outliers on the original input tensor, extract the rolling periodic statistical features using the sliding window method, and generate a multi-source original input dataset with complete spatiotemporal coverage to support the construction of the initial dynamic causal graph.

[0011] Step S2: The multi-source raw input data undergoes time alignment and standardization preprocessing. Based on Granger causality tests and time-series nonlinear dependency analysis, the dynamic driving relationships between variables are identified, generating an initial dynamic causal graph structure. Nodes represent price fluctuation-related factors, and edges represent statistically validated causal directions and strengths. Specifically, this includes: S2.1: Obtain historical price series of agricultural products, climate data, policy announcement texts, supply chain logistics volume and macroeconomic indicators as input datasets. Perform resampling and interpolation processing on the time dimension based on a unified timestamp alignment mechanism to eliminate the heterogeneity of data from different sources in terms of collection frequency and time series granularity, and obtain a time-aligned multi-source time series data matrix to provide a consistent time reference input for subsequent causal analysis. S2.2: Perform zero-mean and Z-score standardization on the time-aligned multi-source time series data matrix to eliminate the numerical bias caused by the difference in the dimensions of each variable, calculate the time series standardization coefficient of each feature, and output the standardized dimensionless time series variable set to ensure that Granger causality test and nonlinear dependency analysis are performed on a fair numerical scale. S2.3: Based on the standardized set of dimensionless time-series variables, the Extended Granger Causality Test is used to perform bidirectional causal significance judgment on each pair of variables. The lagged term coefficients are estimated by the vector autoregression model and the F test is performed to identify statistically significant one-way or two-way causal relationships and generate a preliminary candidate set of causal directions as the input basis for potential directed edges in the dynamic causal graph. For the standardized set of dimensionless time series variables, the extended Granger causality test algorithm (parameters: significance level α is set to 0.05, and the maximum lag order p is determined by the Akaike information criterion) is used to realize the function of determining the two-way causal significance between variable pairs. Furthermore, by constructing a vector autoregressive model VAR(p) (parameter: a state vector containing all target variables, with the order p selected based on the joint minimum of AIC / BIC), the coefficients of different lag terms are estimated, and the significance statistics matrix of each coefficient is obtained; Furthermore, the F-test method (parameters: null hypothesis is that the tested variable does not cause another variable due to Granger causation, alternative hypothesis is that there is a one-way causal effect) is used to determine the overall significance of the coefficients of each group of lagged terms and generate a causal relationship significance scoring table. Furthermore, logical combination processing is performed on the two-way test results of each pair of variables (rule: if both two-way tests are significant, they are marked as two-way causality; if only one direction is significant, they are marked as one-way causality), and the set of causal directions with statistical significance is extracted; A set construction algorithm is used to transform the above set of causal directions into a preliminary candidate set of causal directions, providing input for potential directed edges in the construction of dynamic causal graphs; By using the extended Granger causality test and F-statistic evaluation methods described above, the linear dynamic influence pattern of standardized time series data is transformed into causal candidate data with clear directionality and significance, thereby achieving a reliable identification effect of the initial causal structure. For example, in the scenario of testing the causal relationship between soybean prices and meteorological precipitation, the standardized price series and precipitation series constitute a set of dimensionless time-series variables, and the maximum lag order p is determined to be 4 by the AIC criterion. A two-dimensional VAR(4) model is constructed, and the coefficient matrix is ​​estimated by the least squares method. An F-test is performed on the hypothesis that "soybean prices do not lead to precipitation" to obtain the statistic. The corresponding p-value was 0.018, which is less than the significance level of 0.05, indicating a significant causal relationship between price and precipitation. An F-test was performed on the hypothesis that "precipitation does not affect soybean prices," and the statistical statistic was... The p-value is 0.004, which is also significant, thus establishing a bidirectional causal relationship. This bidirectional action pair is added to the initial candidate set of causal directions and used as an important edge input in subsequent nonlinear dependency analysis and graph structure construction, improving the causal coverage and accuracy of the dynamic causal graph. S2.4: Using the time series mutual information and conditional transfer entropy methods, nonlinear dependency strength analysis is performed on the standardized dimensionless time series variable set to quantify the high-order dynamic influence mode of the superlinear association between variables, extract the nonlinear causal strength matrix, and fuse and weight the matrix with the Granger causality test results to form joint causal determination evidence, which is used to optimize the reliability of causal connections. S2.5: Based on the fused joint causal determination evidence, construct an initial dynamic causal graph structure, map each standardized dimensionless time series variable to a node in the graph, and use the causal relationship that has passed the significance test as a directed edge. The edge weight is determined by the weighted sum of the Granger F statistic and the conditional transition entropy. Output an initial dynamic causal graph containing a set of nodes, a set of directed edges, and dynamic weight parameters, which serves as the basic reasoning framework for subsequent online evolutionary modeling. For the fused joint causal determination evidence set, a node mapping algorithm (parameters: standardized dimensionless time series variable set, node identifier encoding rule) is used to map each time series variable to a single node identifier of the initial dynamic causal graph and record its variable source label and feature dimension index. Furthermore, by using a directed edge generation method (parameters: significance test result set, causal direction candidate set), the causal relationship verified by Granger causality test and nonlinear dependency analysis is transformed into a directed edge structure in the causal graph, and the direction of the edge is determined by the dominant direction label in the causal determination result. Furthermore, the final weight value for each directed edge is obtained through an edge weight fusion calculation method (parameters: Granger F statistic matrix, conditional transition entropy matrix, weighting coefficients α, β). Calculation. The calculation formula is as follows: in, This represents the Granger F-statistic of a given edge. This represents the conditional transition entropy value. and For fusion weighting coefficients; Furthermore, by using a structured encapsulation method for node sets and edge sets, node information, edge information, and dynamic weight parameter sets are organized into the internal data structure of the initial dynamic causal graph, ensuring that they can be directly called and updated in the subsequent online evolutionary modeling stage; By using a causal graph construction algorithm, the results of the previous step are transformed into a graph model containing a set of nodes, a set of directed edges, and weight parameters, thus realizing an initial inference framework that represents the dynamic driving relationship between factors related to price fluctuations. For example, in an agricultural product market analysis scenario, the input variables include the standardized "corn price index". "Precipitation" and "exchange rate" After causality testing, a significant relationship was found: changes in precipitation cause fluctuations in corn prices, and exchange rate changes affect the weighted edges of both corn prices and precipitation. Let the F-value of the effect of precipitation on corn prices in the Granger F-statistic matrix be... The conditional transition entropy value is α take ,β take Then the weights are calculated according to the formula above: = .

[0012] In this scenario, the initial dynamic causal graph consists of three nodes with three directed edges between them. The edge weights correspond to the calculation results mentioned above. The output graph model can be directly updated in the subsequent online evolution stage to reflect new market dynamics, thereby significantly improving the accuracy and robustness of causal relationship representation.

[0013] like Figure 2 As shown, step S3 involves using the initial dynamic causal graph structure as the basic reasoning framework, introducing an online learning mechanism, and adjusting the weight parameters and conditional probability distribution of the causal edges based on real-time updated market feedback data to generate an evolutionary dynamic causal graph model with environmental adaptability, used to characterize the price influence mechanism that changes over time. Specifically, this includes: S3.1: Based on the initial dynamic causal graph structure generated in the previous steps, nodes represent price fluctuation-related factors, and edges represent the causal direction and strength verified by Granger causality test and time-series nonlinear dependency analysis. The sliding time window method is used to extract recent market historical data segments as the input sequence of the online learning mechanism to extract statistical dependency features under the current market conditions. Based on the initial dynamic causal graph structure constructed in the preceding steps, the nodes correspond to price fluctuation-related factors, and the directed edges carry causal direction and dynamic strength parameters obtained through Granger causality test and time-series nonlinear dependency analysis. A sliding time window is used to extract recent market historical data segments as the input sequence of the online learning mechanism to realize continuous updating of causal edge weights and dynamic correction of conditional probability distribution. Using a sliding window selection algorithm (parameter: window length) Depends on price fluctuation cycle and window step size Based on the data refresh frequency, interval truncation is performed on multi-source time series data to ensure that each window contains complete sequences of climate observations, policy events, logistics dynamics, and economic indicators, thereby constructing a sample set of the current market status. Furthermore, by using the time index mapping function, the various types of data streams are aligned according to the correspondence of nodes in the initial dynamic causal graph, forming a node-feature data matrix within the window, which provides structured input for subsequent statistical dependency feature extraction. Furthermore, autocorrelation function analysis is employed (parameter: maximum lag order). Based on the window length setting, the stationarity of each node variable in the window is tested, and the stationary series is processed by variance normalization to eliminate scale differences, ensuring that different factor characteristics can be compared in the same numerical range. Furthermore, by using the mutual information calculation method (parameter: the number of discretized bins B is set based on the sample size), the information transmission strength between any two node variables within the window is quantified to generate a mutual information matrix, providing a nonlinear dependency basis for adjusting the causal edge weights in the online learning stage; Through the above-mentioned sliding truncation and feature dependency analysis processing methods, the node interaction relationship of the initial dynamic causal graph under the recent market state is transformed into a dependency feature set that can be updated online, thereby enhancing the model's real-time perception and adaptability to environmental changes. For example, a statically initialized dynamic cause-effect graph is applied to agricultural market data for the first quarter of 2024, with a set sliding window length. The window duration is 20 trading days. The timeframe is 5 trading days, ensuring each window covers at least one cycle of major climate data updates and policy announcements. In the initial window capture, daily precipitation data from the meteorological bureau, daily closing price fluctuations from the futures exchange, daily logistics inbound volume from customs, and weekly currency exchange rate changes published by the National Bureau of Statistics are mapped to corresponding nodes in a causal graph using time indices, forming a 20×N node feature data matrix (N being the number of factor nodes). Stationarity tests are performed on each column of this matrix, removing original sequences that significantly reject stationarity based on unit root tests, and Z-score normalization is applied to the remaining sequences. Using mutual information calculation, with a discretization bin number of B=10, a mutual information matrix between node pairs is obtained. Non-zero terms in the matrix represent information dependencies within the window. For example, a significant increase in the mutual information value between the node "precipitation" and the node "spot price volatility" reflects an enhanced driving effect of climate variables on price fluctuations during this time interval. Thus, the resulting dependency feature set is input into the subsequent incremental Granger causality strength re-estimation module to realize the dynamic adjustment of causal edge weights, thereby constructing an evolutionary dynamic causal graph input basis that is more in line with the current actual market state. S3.2: Perform incremental Granger causality strength re-estimation on the multi-source time series data within the sliding time window, and use recursive least squares (RLS) to dynamically update the directional strength coefficient of each causal edge, and calculate the corrected causal influence metric matrix as the initial adjustment basis for the edge weight parameters in the evolutionary dynamic causal graph. S3.3: Based on real-time market feedback data, including the fluctuation range of spot prices of agricultural products, abnormal trading volume of futures contracts, policy release event markers and logistics interruption alarm signals, a high-dimensional observation vector is constructed and input into the Bayesian variational inference module to estimate the changing trend of conditional probability distribution on each causal path. S3.4: The modified causal influence measurement matrix and the conditional probability distribution change results obtained by Bayes inference are fused together, and a joint confidence score is generated using a weighted adaptive fusion algorithm. Based on this, the comprehensive weight parameters and node transition probabilities of each edge in the evolutionary dynamic causal graph are optimized to form a graph model parameter set that reflects the latest market-driven mechanism. Sub-step S3.4 is placed in the logical position of main step S3. Its input conditions include the corrected causal influence measurement matrix output by S3.2 and the conditional probability distribution change trend dataset estimated by the Bayesian variational inference module output by S3.3. A weighted adaptive fusion algorithm (parameters: initial weight coefficients are derived from historical prediction accuracy, and the fusion learning rate η is set to 0.01) is used to jointly calculate the modified causal influence measurement matrix and the conditional probability distribution change vector to obtain a preliminary fusion matrix. Furthermore, through normalization (method: extreme value normalization, mapping each column of data to the [0,1] interval), the scale of data with different dimensions is unified, and a normalized fusion matrix is ​​obtained; Furthermore, using the confidence score calculation formula (formula structure: joint weight multiplied by conditional probability correction coefficient), a joint confidence score is generated for each directed edge: in, For joint confidence scoring, To integrate the subsequent weights, This represents the trend value of the conditional probability distribution. Furthermore, adaptive thresholding is performed on the aforementioned confidence score set (parameter: threshold). Based on the mean and variance of the confidence score, the edge set is filtered to obtain a list of high-confidence causal connections; Furthermore, the joint confidence score and the transition probability matrix between nodes are fused using matrix multiplication and addition to calculate the updated node transition probabilities: in The updated node transition probability. The original node transition probability is used. Through this fusion process, the results of the previous step are transformed into a globally consistent graph model parameter set, realizing the optimization of edge weights and node transition probabilities in an evolutionary dynamic causal graph that reflects the latest market-driven mechanisms. For example, in the agricultural futures market scenario, the adjusted weight of a certain edge of the causal influence measurement matrix (reflecting the impact of climate anomalies on spot prices) is 0.62, and the trend value of the conditional probability distribution is calculated to be 0.75. The joint confidence score is then calculated using the above formula: get During dynamic thresholding, the system sets the threshold to 0.40, and this edge is retained as a high-confidence connection. When fusing node transition probabilities, the original probability is... Update calculation as get This significantly improves the activity of the path during the inference process. This embodiment shows that the fused graph model can more accurately characterize the driving effect of climate anomalies on prices in the current time window, and improves the accuracy and robustness of the strategy generation module in the face of sudden climate events; S3.5: Perform sparsity regularization processing on the evolutionary dynamic causal graph model, remove weak causal connection edges with confidence below a set threshold based on the dynamic threshold pruning strategy, while retaining significantly enhanced emerging causal paths, and output an updated dynamic causal graph with a simplified structure and focused semantics, as the shared inference framework of the next stage dual-channel graph neural network architecture. Based on the parameter set of the evolutionary dynamic causal graph model optimized by the weighted adaptive fusion algorithm, the sparsification regularization method with L1 norm constraint (parameter: λ is the sparsification intensity coefficient) is used to achieve sparsification of the causal edge weight matrix, and the edge weights that do not have significant statistical contributions converge to zero. Furthermore, a connectivity elimination operation is performed on the sparsified causal edge set using a dynamic threshold pruning strategy (parameter: threshold θ is adaptively adjusted according to real-time market volatility) to remove weak causal connections with joint confidence scores below θ, thereby eliminating noise interference to the graph structure inference performance. Furthermore, a causal emerging detection algorithm (parameters: window length W and enhancement rate η) is used to calculate causal paths with a confidence enhancement rate exceeding η within the sliding window W and mark them as emerging paths, ensuring that these paths are preserved during the pruning process to reflect the breakthrough of the latest market drivers. Furthermore, through graph structure reconstruction, the significantly enhanced emerging causal paths and high-confidence existing paths are recombined to generate a simplified and semantically focused causal graph structure, and the node adjacency table and edge weight index matrix are updated to ensure the consistency and efficiency of subsequent dual-channel graph neural network inputs. By combining sparsification regularization with dynamic pruning, the high-dimensional causal graph model result from the previous step is transformed into a dynamic causal graph structure with structural compression and reasoning centralization, thereby achieving the expected technical effect of reducing computational complexity and enhancing path interpretability. For example, in agricultural product price hedging applications, the edge weight matrix of an evolutionary dynamic causal graph model is input into an L1 regularization module, with λ set to 0.05. After sparsification, a large number of weight values ​​converge to... The following applies dynamic threshold pruning to the sparse matrix, where the threshold θ is based on market volatility. Calculated as In recent volatility In the case that θ is Remove all edges with an edge confidence level below a certain value. An emerging potential detection algorithm with a window length W = 15 days and an enhancement rate η = 1.2 is used. Before pruning, the edge confidence changes over the past 15 days are analyzed, and paths with an enhancement rate ≥ 1.2 are retained, such as "Southern Drought → Reduced Production Expectations → Spot Price Volatility". The edge weights of this path are determined by... Growth to The reconstructed causal graph reduces the number of nodes to 70% and the number of directed edges to 65%, significantly shortening the inference time of the subsequent dual-channel graph neural network. At the same time, the relevance of the core driving path in the explanation module is significantly improved, enhancing the transparency and acceptability of the overall policy explanation.

[0014] like Figure 3As shown, step S4 involves constructing a dual-channel graph neural network architecture based on the evolutionary dynamic causal graph model. The policy generation channel and the explanation generation channel share the same graph structure representation, and a synchronous attention gating mechanism ensures that the two channels maintain state consistency in each inference step, forming a joint decision-explanation inference space. Specifically, this includes: S4.1: Based on the set of nodes and the set of weighted directed edges in the evolutionary dynamic causal graph model, construct the underlying topology of the graph neural network, where each node corresponds to a price influence factor and each edge represents the causal strength weight and conditional probability distribution updated through online learning, so as to form a graph-structured representation input that supports temporal reasoning. Based on the node set and weighted directed edge set output by the evolutionary dynamic causal graph model after online learning and sparsification in the S3 step, a graph structure construction algorithm (parameters: node identifier list, edge start and end index, edge weight and conditional probability matrix) is used to map price fluctuation related factors to node entities at the bottom layer of the graph neural network. Furthermore, by using the node attribute initialization method (parameters: multi-source time-series feature vector, normalized coefficient set), the economic and meteorological characteristics, policy event labels, and logistics status of each node are encoded into an initial feature vector of fixed dimensions, and the node feature matrix is ​​obtained. Furthermore, an adjacency matrix generation algorithm (parameters: directed edge set, edge weight, conditional probability distribution) is adopted to transform causal edges into a numerical adjacency matrix structure, and to ensure that the edge weight and conditional probability are embedded in the corresponding positions of the matrix in a dual-channel manner, thereby generating graph connection information that supports temporal reasoning. Furthermore, by utilizing the temporal causal embedding method (parameters: node feature matrix, adjacency matrix, time step index), a graph-structured representation input that can explicitly represent the dynamic relationship between price drivers is constructed, generating a unified topological framework for strategy generation and interpretation of the generation channel; Furthermore, a topology consistency verification algorithm (parameters: node index set, adjacency matrix, graph connectivity threshold) is adopted to verify the connectivity and directionality of the generated underlying topology, remove isolated nodes or invalid causal edges, and form a graph model input with valid structure and accurate semantics. Through the above graph structure construction and feature mapping processing, the evolutionary dynamic causal graph result from the previous step is transformed into structured data that can be directly parsed by a dual-channel graph neural network, achieving a unified and stable input effect of the underlying topology in policy and interpretation reasoning; For example, in an agricultural product price hedging task, the evolutionary dynamic causal graph model, after the S3.5 sub-step, outputs 120 nodes and 420 directed edges. The node feature dimension is set to 64, including normalized meteorological indicators, macroeconomic vectors, and event trigger markers. The graph structure construction algorithm inputs node indices 0 to 119, corresponding to price fluctuation-related factors such as "rainfall in the south," "changes in futures positions," and "policy tariff adjustments." The adjacency matrix is ​​generated from the 420 directed edges, with edge weights ranging from 0.15 to 0.92. The conditional probability distribution matrix and the edge weight matrix are embedded synchronously according to row and column indices. Using a temporal causal embedding method, the node feature matrix and the adjacency matrix are embedded within time step t=30 to obtain a unified topological framework tensor. , dimension During the topology consistency verification process, 12 edges with confidence scores below 0.1 were removed, and the final topology's shortest path length distribution under connectivity detection met the preset threshold condition. This input structure significantly improved the consistency of node importance assessment and the synchronization of policy-interpretation during the subsequent dual-channel graph neural network initialization process. Verification results show that it can maintain the logical stability and interpretability of decision outputs under multiple market state samples. S4.2: Based on the graph structured representation, initialize the dual-channel graph neural network architecture: the policy generation channel uses a graph attention network (GAT) to aggregate and calculate node features and extract key state embedding vectors for action decision-making; the interpretation generation channel deploys GAT encoders with the same structure in parallel, sharing underlying parameters and adjacency relationships to ensure the consistency of semantic understanding of the same causal graph by the two channels. Based on the node set and weighted directed edge set of the evolutionary dynamic causal graph as input, the encoder structure of the channel is generated by the graph attention network (GAT) initialization strategy. The number of attention layers is set to three, each containing eight parallel attention heads. The activation function is LeakyReLU (parameter: negative slope 0.2) to realize nonlinear weighted aggregation calculation of node features. Furthermore, by performing layer-by-layer weighted summation and normalization on the adjacency matrix and node feature matrix of the evolutionary dynamic causal graph, and using an adaptive attention weight update mechanism (parameter: learning rate 0.001), key state embedding vectors reflecting the correlation strength of price driving factors are extracted and used as decision input features for the policy generation channel. Furthermore, a graph attention encoder structure for interpreting the generated channels is constructed, the network topology parameters and adjacency matrix of the policy-generated channels are reused, a shared weight mechanism is used to lock the underlying parameters to ensure semantic consistency, and the attention weight calculation method of the same as that of the policy channels is used to ensure that the attention distribution of the two channels is consistent on the same graph structure. Furthermore, a parameter synchronization update algorithm is adopted to perform stepwise averaging and difference constraints on the weight gradients obtained by the two channels in the training iteration, and the consistency loss function is used to adjust the weights of both sides, so that the policy generation channel and the interpretation generation channel maintain a high degree of alignment in node importance judgment and feature aggregation results. By using a dual-channel shared underlying topology and parameter initialization method, the structural information of the evolutionary dynamic causal graph is transformed into two sets of state embedding vectors with different functions but consistent semantics, which serve policy action calculation and causal path extraction respectively, achieving the initial consistency technical effect of policy recommendation and interpretation generation. For example, in a specific agricultural product market scenario, the node set size is 50, representing 50 price influencing factors; the adjacency matrix sparsity is 0.15, and the edge weights range from 0.05 to 0.9. The first layer of the GAT in the policy generation channel has 8 attention heads, each with a feature dimension of 64, and the input node feature matrix has a dimension of 50×32. Attention calculation is then performed. In the interpretation generation channel, the attention head weight matrix and attention coefficients of the policy generation channel are directly referenced to ensure complete consistency in the output results of the function σ between the two channels. The actual execution effect shows that the cosine similarity of the embedding vectors of the policy generation channel and the interpretation generation channel reaches 0.99, significantly improving the consistency of the state semantics between the two channels and providing a stable foundation for the subsequent implementation of the synchronous attention gating mechanism. S4.3: Design a synchronous attention gating mechanism to compare the attention weight matrix of each inference step in the policy generation channel with the corresponding matrix in the interpretation generation channel in real time, and constrain the difference between the two to not exceed a preset threshold through a consistency loss function, thereby ensuring that the two channels remain synchronized in node importance assessment and path activation selection, forming a joint reasoning process with consistent causal perception. After the initial structure and topological relationship of the dual-channel graph neural network are loaded, a real-time monitoring mechanism for attention weights (parameters: time step index, node feature tensor, adjacency matrix) is adopted to realize the function of capturing and interpreting the attention distribution of the policy generation channel in each inference step and synchronously reading the corresponding distribution of the generation channel. Furthermore, the attention matrix difference calculation method (parameter: policy channel attention matrix) is used. Explain the channel attention matrix This allows for element-wise quantification of the differences between the two channels in node importance scoring, resulting in a difference matrix. ,in ; Furthermore, by constructing a consistency loss function (parameters: preset difference threshold τ, consistency coefficient λ), the difference matrix is ​​optimized. Perform norm calculations to generate loss values. This is used to constrain the weight distribution difference to not exceed τ, ensuring the consistency of node activation selection between channels; The formula is as follows: in, λ represents the Frobenius norm of the dissimilarity matrix, and λ is the consistency constraint coefficient. Furthermore, a consensus optimization method based on gradient backpropagation (parameters: learning rate η, optimization rounds T) is adopted to reduce the consensus loss. The parameters are passed back to the two-channel graph attention network parameter space, and the distribution of node weights and edge weights is gradually adjusted so that the norm of the difference matrix converges to the threshold range. By using a synchronous attention gating mechanism, the difference suppression result of the previous step is transformed into a causal path activation mask matrix, achieving complete synchronization of the two channels in path selection and ensuring consistent causal perception in the joint reasoning process. For example, in a certain agricultural futures hedging scenario, the attention weight matrix of the strategy generation channel at time step t The dimension is 50×50, which explains the attention weight matrix of the generation channel. Same-dimensional matching, preset difference threshold τ is set to Let the consistency coefficient λ be set as Dissimilarity matrix Calculated using the Frobenius norm Consistency loss value for = In the optimization round T = In the iteration, the learning rate η is The norm of the dissimilarity matrix decreased to [value missing] in the 10th round. Below the threshold τ, node importance and path activation are synchronized between the two channels. The final output mask matrix is ​​50×50, activating only strongly correlated causal paths, significantly improving the consistency and reliability of the policy output and interpretation results; S4.4: Based on the synchronized attention distribution, the optimal action probability distribution is calculated in the policy generation channel, and the set of core causal paths that are significantly activated is identified in the interpretation generation channel. Both are based on the evolutionary dynamic causal graph model to achieve the simultaneous generation of policy action suggestions and potential interpretation paths. Under the constraint of synchronous attention gating mechanism, the input is the attention weight distribution matrix of the policy generation channel and the interpretation generation channel at the same inference time step, and the node feature vector sequence of the evolutionary dynamic causal graph is used as the basis for joint computation. The algorithm for estimating the probability of action in a graph attention network in the strategy generation channel is adopted (parameters: node embedding dimension d, number of attention heads h, negative slope of LeakyReLU 0.2). After weighted aggregation of node features in the normalized attention distribution matrix, the state-action probability distribution is calculated to realize the probability quantification evaluation of each potential hedging operation in the current environment. Furthermore, by interpreting the causal path activation recognition algorithm of the generated channel (parameters: activation threshold τ, path length upper limit Lmax), high-weight path tracking is performed on the synchronized attention distribution to identify the set of continuous directed edges with weights greater than τ in the evolutionary dynamic causal graph, and to generate the core causal path candidate set. Furthermore, using the path importance calculation formula: in, For path importance, Let be the attention weight of the i-th edge of the path. This represents the overall weight of the edge in the evolutionary dynamic causal graph. The path length is used to quantify the contribution of the path to the current strategy action, and to select the set of the top K core causal paths by importance. Furthermore, the probability distribution of policy actions is bound to the core causal path set through a joint mapping mechanism to form an action-explanation dual metadata structure, which prepares a structured input for the subsequent decision fusion module. Through the above algorithm chain, the synchronous attention distribution is transformed into the action probability distribution of the policy generation channel and the core causal path set of the explanation generation channel, so as to achieve the synchronous output of the two in terms of time step and causal logic, and ensure the strict matching between policy and explanation. For example, with the input node embedding dimension d=128, attention head number h=8, activation threshold τ=0.75, and path length upper limit Lmax=5, the strategy generation channel outputs the probability value of "short" futures position opening direction under the current market conditions. Hedging probability using option combinations Probability of spot inventory rebalancing The explanation states that the generation channel traces the core causal path "abnormal southern climate → decreased production expectations → increased spot price volatility → short-selling futures strategy" at the same time step. The path importance formula is used to calculate the side attention weights, which are as follows: , , The edge comprehensive weights are respectively , , The path length is 4, and the path importance is obtained. ≈ This path enters the core set and is bound to the short-selling strategy probability value. Finally, the results generated simultaneously by the two channels are verified for causal consistency in the decision fusion module. Market backtesting shows that the strategy adoption rate explained by the binding is significantly improved, and the volatility of execution returns is significantly reduced. S4.5: The state-action value function output by the policy generation channel and the causal path activity score output by the explanation generation channel are jointly input into the decision fusion module to construct a joint decision-explanation inference space, which serves as a unified semantic benchmark framework for subsequent reinforcement learning search and semantic mapping operations. Based on the state-action value function obtained from the policy generation channel and the causal path activity score output from the explanation generation channel, a vector-level normalization method (parameter: L2 norm) is used to achieve consistent processing of data from different sources on numerical scales. Furthermore, a weighted fusion algorithm (parameters: weight coefficient α based on the historical performance score of the strategy channel, weight coefficient β based on the path contribution score of the explanation channel) is used to achieve a weighted combination of the output results of the two channels and obtain the fused joint feature vector set. Furthermore, a covariance matching method is adopted (parameter: the expected covariance matrix is ​​calculated from the correspondence between the historical best strategy and the interpretation) to adjust the correlation of the joint feature vector and generate a joint decision feature matrix optimized by correlation. Furthermore, a unified semantic space mapping of policy and explanatory features is achieved through a multimodal embedding mapping algorithm (parameters: embedding dimension is set to 128, loss function is joint cosine similarity minimization), and semantic embedding vectors for subsequent reasoning are generated. Through tensor concatenation operations, the semantic embedding vectors mentioned above are transformed into indexed representations of the joint decision-explanation reasoning space, enabling co-location storage of policy actions and explanation paths in a unified model coordinate system, and providing a consistent semantic benchmark for reinforcement learning search and semantic mapping operations. For example, in the process of generating agricultural product hedging strategies, the state-action value function output by the strategy generation channel is a set of real-valued vectors of length 64, and the causal path activity score output by the interpretation generation channel is a set of normalized scores of length 64. Both are normalized using the L2 norm, resulting in two vectors of unit length. Weighting coefficients α=0.6 and β=0.4 are then used for weighted fusion to generate a joint feature vector. For instance, if the 10th dimension of the state-action value function is 0.85 and the 10th dimension of the causal path score is 0.65, the calculated weighted fusion value is... =0.77. A covariance matching method was used to adjust the fused vector to be close to the expected covariance matrix of the historical best policy-explanation pair. Finally, an embedding mapping algorithm was used to map the 64-dimensional joint features to a 128-dimensional semantic space, and the cosine similarity loss value converged to 0.02. The indexed joint features generated by the above processing were used for subsequent reinforcement learning search, realizing a unified inference starting point for policy and explanation, and significantly improving the stability of policy recommendation and explanation consistency on multiple batches of real-time market data.

[0015] Step S5: In the joint decision-explanation reasoning space, a reinforcement learning algorithm is used to perform an optimal hedging action search with the objective function of minimizing expected price risk, generating a strategy operation sequence, and simultaneously recording the set of core causal paths activated during this process. Specifically, this includes: S5.1: The node embedding vector sequence output by the evolutionary dynamic causal graph model is used as the state input of the graph neural network. The attention mechanism is used to weight and aggregate the various influencing factors to calculate the joint representation vector of the current market state, so as to capture the nonlinear interaction characteristics between multidimensional variables and form the initial decision basis for the strategy generation channel. Based on the node embedding vector sequence output by the evolutionary dynamic causal graph model, a graph attention mechanism (parameters: adjacency matrix, node feature matrix, attention weight initialization value) is used to perform weighted aggregation calculation on the price influence factors of each node, so as to realize the adaptive adjustment of the weights between nodes and highlight the factor nodes that contribute more to the current market state. Furthermore, a multi-head graph attention network method is used (parameters: number of parallel attention heads k, feature dimension of each head). This enables the aggregation of factor information from multiple perspectives and yields a high-dimensional node representation matrix that integrates causal information from multiple paths, thereby enhancing the ability to capture nonlinear interactions. Furthermore, a temporal context fusion algorithm (parameters: sliding time window length w, decay coefficient λ) is used to perform temporal weighted combination of the node representations after attention aggregation, forming a temporal embedding matrix that considers the causal time delay effect, and generating a dynamic feature set that reflects the current market evolution trend. Furthermore, through feature normalization processing methods (parameter: normalization range [...]), [1,1], standardization of mean μ and variance σ) to achieve scale unification of features with different dimensions, and output dimensionless normalized embedding vectors to reduce the computational bias of subsequent policy generation algorithms; Furthermore, a feature fusion mapping function F is employed to integrate the normalized embedding vectors into a joint representation vector through a nonlinear activation transformation, thereby characterizing the nonlinear interaction patterns among multidimensional market driving factors. This mapping function... The format is as follows: in, For a normalized set of embedded vectors, For the mapping weight matrix, For bias terms, It is a non-linear activation function (such as ReLU); Through the aforementioned attention aggregation, multi-view fusion, temporal context calculation, normalization, and nonlinear mapping processing, the embedded sequence result of the previous step is transformed into a joint representation vector, thereby enhancing the input state's nonlinear perception of the current market and providing an initial decision basis for the strategy generation channel. For example, in a soybean futures market application scenario, the evolutionary dynamic causal graph model has 50 nodes, including factors such as climate, logistics, policy, and macroeconomic indicators. When using the graph attention mechanism, the adjacency matrix sparsity is set to 0.2, and the attention weights are initialized to a uniform distribution U(0,0.1). The multi-head attention network has 8 parallel heads, each with a feature dimension of 16. The temporal fusion sliding window length is 15 trading days, and the decay coefficient λ is set to 0.85. The normalization parameters are selected with mean μ=0 and variance σ=1, and the normalization range is set to... 1 to 1. After the above processing, the joint representation vector obtained by the node embedding vector calculation has a dimension of 128. In the state input stage of the subsequent reinforcement learning strategy generation process, it shows a significant improvement in feature sensitivity and causal path correlation. Its response speed to sudden market drought and policy-related export restriction events is significantly faster, the risk control effect of decision-making actions is improved, and the final output hedging scheme achieves stable risk exposure suppression during the backtesting period. S5.2: Input the joint representation vector into the reinforcement learning framework built on deep deterministic policy gradient (DDPG), perform action space mapping, and generate a continuous hedging operation suggestion sequence as a strategy operation sequence, including futures position building direction, option combination allocation ratio and spot position adjustment range, so as to minimize the expected risk loss caused by long-term price fluctuations; The joint representation vector is input into a reinforcement learning framework based on Deep Deterministic Policy Gradient (DDPG), and an Actor-Critic dual-network structure is adopted (Actor network parameters: 128 hidden nodes, ReLU activation function; Critic network parameters: 256 hidden nodes, ReLU activation function) to realize the mapping function from high-dimensional state vector to continuous action space. Furthermore, using the deterministic policy gradient calculation method in the policy update module, based on the state-action value function output by the Critic network... Gradient backpropagation is performed on the Actor parameters to ensure that the generated action suggestions minimize price risk under the long-term expected return. Furthermore, the target network's delayed update mechanism (parameter: soft update coefficient) is utilized. Pick In each iteration, the main network parameters are synchronized to the target network in a weighted manner to improve the stability of the training process and reduce policy oscillations. Furthermore, through the experience replay buffer (capacity: (Each state transition record) performs a batch sampling operation, randomly inputting past state transition samples into the Actor-Critic network to break the temporal correlation between samples and improve the network's generalization ability; Furthermore, based on the optimization objective function ,in For price risk reward, The mean, Using the sample size, risk-sensitive screening and adjustment are performed on the generated continuous action vectors to ensure that the direction of position building, the proportion of option combinations, and the magnitude of spot position adjustment are optimally allocated under the constraint of minimizing expected risk. Through the reinforcement learning algorithm module and parameter constraint processing method described above, the joint representation vector of the previous step is transformed into a continuous hedging operation suggestion sequence containing futures position building direction, option ratio allocation and spot position adjustment range, thereby achieving the technical effect of reducing the expected risk loss caused by price fluctuations. For example, in an agricultural futures hedging scenario, the state input is a 48-dimensional joint representation vector containing weather-driven factors, supply chain disruption indices, and macroeconomic pressure indices. The Actor network is set with an input layer of 48 nodes, two hidden layers of 128 and 64 nodes respectively, and the output action vector contains three dimensions, corresponding to the futures position opening direction (value range). Option portfolio ratio (range of values) and the magnitude of spot inventory adjustments (range of values) The Critic network takes a joint state-action vector as input, has hidden layers of 256 and 128 nodes respectively, and outputs a single risk-reward value. The experience replay buffer capacity is set to... 1 record, sampling batch size is During training, the soft update coefficients... Discount factor used for weight updates at each step This is applied to future return calculations. The resulting continuous hedging operation recommendations guide futures position opening. (Represents increasing short positions), option portfolio ratio (Represents increasing the weighting of protective call options), spot position adjustment range (This represents a 20% reduction in spot holdings), and the verification results show that the portfolio's price volatility risk has been significantly reduced in the coming quarters. S5.3: In each decision time step, the interpretation generation channel is activated synchronously, the set of causal edges that are significantly activated during the policy generation process is tracked, the key driving path is locked based on the synchronous attention gating mechanism, and the corresponding intermediate causal trajectory sequence is generated as the original input of interpretability information. S5.4: Perform path pruning and importance ranking on the intermediate causal trajectory sequence, use the causal strength threshold to filter out the core causal path set that contributes more to the current strategy action than a preset standard, and mark it as an explanatory support chain that is strongly related to this hedging operation; S5.5: The generated policy operation sequence and its corresponding core causal path set are used as a joint output result. Before being passed to the next verification module, time alignment processing is performed to ensure that each action step and its explanation path are strictly matched in terms of timestamp and causal logic, forming a closed-loop consistent policy-explanation dual output flow.

[0016] Step S6: Input the core causal path set into the semantic mapping module, and convert it into a logical chain description readable in natural language based on a predefined causal semantic rule base, generating interpretable text fragments that strictly correspond to the policy actions, thus forming the policy explanation output. Specifically, this includes: S6.1: Obtain the set of core causal paths recorded and output by the dual-channel graph neural network during the strategy generation process. Each causal path represents the activation inference link from a certain external influencing factor (such as climate anomaly, policy change) through several intermediate nodes (such as supply expectation, market sentiment) in the dynamic causal graph to the final hedging action (such as increasing short position), and serves as the input object for semantic mapping. S6.2: Based on the node semantic labels and edge causal relationship types in the core causal path set, use the mapping templates stored in the pre-built causal semantic rule library to perform pattern matching and identify the natural language expression structure corresponding to each path segment. For example, map 'drought → reduced production expectation ↑ → spot price volatility ↑' to "due to the continuous drought in the main producing areas, the expected reduction in crop production has increased, which in turn has pushed up the volatility of spot market prices". Based on the node semantic labels and edge causal relationship types in the core causal path set, a causal template pattern matching algorithm (parameters: node label set, edge type encoding, rule base index key) is used to realize the binding and matching function between path elements and semantic templates; Furthermore, by using the semantic rule base retrieval function (parameters: sequence of node labels in the path, sequence of causal relationship types), the path nodes are bound one-to-one with the rule template slots, and the original set of natural language fragments is obtained. Furthermore, a semantic conflict resolution algorithm (parameters: rule template group, thesaurus, conflict detection threshold) is adopted to realize the conflict detection and replacement function between different templates and generate a set of semantically consistent fragments; Furthermore, through a structured semantic fusion method (parameters: matching fragment set, node activation weight matrix), the fragments are arranged in logical order and linguistically connected in causal relationships, and a paragraph-level natural language expression structure with temporal logic is generated. Through the above pattern matching and fusion processing, the core causal path set of the previous step is transformed into itemized natural language expression paragraphs, realizing the readable and standardized output effect of the causal chain; For example, in the process of generating an agricultural product price hedging scheme, the core causal path set of the input contains three node sequences: "abnormal climate → increased expectation of reduced production → increased volatility in spot prices", "reduction in policy subsidies → decrease in planting area → decline in output", and "depreciation of international exchange rate → decrease in export profits → pressure on futures prices". The rule base pre-sets template T1 as "due to climate events, changes in expected production lead to changes in price volatility", T2 as "due to policy events, changes in supply lead to changes in output", and T3 as "affected by exchange rate events, changes in export profits lead to changes in futures prices". The pattern matching algorithm sequentially binds the first path node labels {abnormal climate, increased production reduction expectations, increased spot price volatility} to the T1 template slot, generating the initial statement "Due to abnormal climate in major crop-producing areas, the expectation of reduced production has increased, which in turn has pushed up spot price volatility." The second path is bound to T2, generating "Due to reduced government subsidies, the planting area has decreased, ultimately leading to a decline in output." The third path is bound to T3, generating "Affected by the decline in international exchange rates, export profits have decreased, which in turn has put pressure on futures prices." During the semantic conflict resolution stage, a synonym for "decline" is detected in the second and third explanations. The algorithm replaces "decline" with "weaken" in the third sentence, generating "Affected by the decline in international exchange rates, export profits have weakened, which in turn has put pressure on futures prices." The semantic fusion method arranges sentences according to the node activation weight matrix and connects them with temporal logical connectors to form a paragraph structure: "Due to abnormal weather in major crop-producing areas, expectations of reduced production have increased, which in turn has pushed up the volatility of spot prices. At the same time, reduced government subsidies have led to a decrease in planting area, ultimately resulting in a decline in output. In addition, the decline in international exchange rates has weakened export profits, which in turn has put pressure on futures prices." This paragraph, as an explanatory output, is linked to corresponding hedging actions (such as increasing short positions in futures), significantly improving the understandability and credibility of the strategy recommendations. S6.3: Perform temporal logic integration and grammatical optimization on the matched natural language expression structure, adopt a sentence fusion algorithm based on dependency parsing to merge multiple short sentences into coherent complex sentences, and introduce a referential resolution mechanism to eliminate repeated subjects or redundant expressions, generating a grammatically fluent and logically clear complete explanatory text paragraph. S6.4: The generated natural language explanation text is spatiotemporally aligned with the original hedging operation instructions. Based on the time step index and key node activation weight recorded by the synchronous attention gating mechanism, it is ensured that each explanation content accurately corresponds to the decision motivation at a specific time point in the strategy sequence, forming a traceable explanation output with strict action-basis binding. S6.5: Output a structured, interpretable text fragment containing a complete logical chain of hedging action identifiers, triggering conditions, core driving factors, intermediate transmission paths, and final impact inferences, as input to the subsequent consistency verification module.

[0017] Step S7: Determine whether the logical consistency threshold requirement is met between the strategy operation sequence and the corresponding interpretable text segment. If not, trigger the causal subgraph recalibration mechanism to re-retrieve or optimize relevant causal paths based on the latest market state until consistency is reached. Specifically, this includes: S7.1: Based on the strategy operation sequence and the corresponding interpretable text fragments, extract the set of core causal paths on which the strategy actions depend, and use them as the input basis for logical consistency comparison to form a structured decision tracing chain; S7.2: Perform semantic alignment processing on the core causal path set and the natural language explanation text, and use the semantic role labeling model based on dynamic causal graph to identify the factor nodes, causal edge directions and conditional probability change trends mentioned in the explanation text, and generate a structured analytical graph. S7.3: Calculate the topological similarity between the structured analytical graph and the causal subgraph actually activated in the policy generation channel, and use the graph editing distance algorithm to quantify the degree of deviation between the two in terms of node connectivity, causal directionality and weight distribution to obtain a logical consistency score; S7.4: Determine whether the logical consistency score meets the preset threshold requirement. If it is lower than the threshold, it is determined that there is a strategy-interpretation logical gap, triggering the activation condition of the causal subgraph recalibration mechanism. Based on the logical consistency score data output from step S7.3, a threshold determination algorithm is used (parameter: preset logical consistency threshold). This enables a quantitative comparison of the matching degree between the strategy operation sequence and the explanatory text fragment; Furthermore, by comparing logical consistency scores and The numerical relationship is determined using the judgment condition formula: To achieve a consistent state identifier, where The variable is used to identify the result; a value of 1 indicates consistency, and a value of 0 indicates inconsistency. Furthermore, when determining the outcome variable When the value is 0, the strategy-interpretation consistency fault identification logic marks the risk of a causal chain break in this hedging scheme and generates a triggering signal. ; Furthermore, by triggering the identification signal The startup interface of the causal subgraph recalibration mechanism is called to realize the subsequent re-retrieval of causal paths and preparation for structural parameter optimization. Through this judgment and triggering process, the logical consistency score result of the previous step is transformed into a clear mechanism activation condition, so as to achieve the expected technical effect of rapid identification and intelligent repair of policy-interpretation logical gaps. For example, in a certain calculation of an agricultural futures hedging strategy, the topological similarity score between the structured parsed graph and the actual activated causal subgraph is: Preset logical consistency threshold for Substitute the criteria into the formula: The judgment result is obtained. The system generates accordingly. The trigger signal indicates insufficient logical consistency, requiring the execution of the causal subgraph recalibration mechanism. After initiating the recalibration mechanism, the retrieved causal subgraph paths show a significant improvement in node connectivity and weight distribution, aligning well with the policy actions, ultimately achieving a consistency score in the re-evaluation. If the threshold is exceeded, the logical closed-loop consistency of the solution is restored. S7.5: Based on the latest market feedback data and the evolutionary dynamic causal graph model updated through online learning, the most relevant causal subgraph path to the current strategy action is re-examined, and its structure and parameters are optimized through reverse reasoning to generate a set of consistency-enhanced causal paths that match the strategy action, providing a reliable basis for subsequent reinterpretation.

[0018] Step S8: Output the final hedging scheme, which includes hedging operation instructions that have been verified for consistency and their complete decision-making basis chain, realizing a closed-loop collaborative output of strategy recommendation and interpretation generation under a unified model. Specifically, it includes: S8.1: Based on the hedging operation sequence that has been verified by logical consistency and the corresponding interpretable text fragments, construct a structured hedging scheme data package, which includes operation type, underlying asset, execution sequence, position direction, risk exposure adjustment amount and related causal path identifier, to form a standardized output carrier; S8.2: Perform integrity verification on the structured hedging scheme data packet, check whether each field meets the preset syntax specifications and business constraints. If missing or conflicting items are found, trigger the exception handling mechanism and fall back to the cause-effect subgraph recalibration process in S7 to ensure the compliance and consistency of the output content. S8.3: The structured data packets that have passed integrity verification are input into the multimodal encapsulation module, which converts them into a dual-channel output format of visual report format and machine-readable interface format based on a predefined template engine. The former is used to support user terminal display, and the latter is used to connect to the automated trading system to execute instructions. S8.4: Digital signature and timestamp information are embedded synchronously during the encapsulation process. Asymmetric encryption algorithms are used to perform security authentication on the output scheme and generate a unique traceability certificate to prevent the content of the scheme from being tampered with and to support subsequent audit tracing, thereby improving the security and legal compliance of the system. S8.5: The final hedging solution, after being packaged and certified, is pushed to the user interface and the background execution queue, respectively for manual review and confirmation and for automatic trading system to call, achieving seamless connection from intelligent decision-making to practical application and completing the terminal output function of the entire closed-loop system.

[0019] 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.

[0020] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for hedging agricultural product price data using electronic data processing, characterized in that, Includes the following steps: S1: Obtain historical price series of agricultural products, climate data, policy announcements, supply chain logistics information, and macroeconomic indicators as multi-source raw input data; S2: Perform time alignment and standardization preprocessing on the multi-source raw input data, identify the dynamic driving relationship between variables, and generate an initial dynamic causal graph structure; S3: Using the initial dynamic causal graph structure as the basic reasoning framework, an online learning mechanism is introduced to adjust the weight parameters and conditional probability distribution of the causal edges based on real-time updated market feedback data, thereby generating an evolutionary dynamic causal graph model. S4: Based on the aforementioned evolutionary dynamic causal graph model, a dual-channel graph neural network architecture is constructed, and a synchronous attention gating mechanism is introduced to maintain the consistency of the state of the two channels in each reasoning step, forming a joint decision-making and explanatory reasoning space; S5: In the joint decision-making and explanation reasoning space, with the objective function of minimizing expected price risk, perform the optimal hedging action search, generate a strategy operation sequence, and record the set of core causal paths activated during the optimal hedging action search process; S6: Perform semantic mapping on the core causal path set, convert it into a logical chain description readable in natural language based on a predefined causal semantic rule base, and generate interpretable text fragments corresponding to the policy actions.

2. The method for hedging agricultural product price data through electronic data processing according to claim 1, characterized in that, Following step S6, the following is also included: S7: When the logical consistency threshold requirement between the strategy operation sequence and the corresponding interpretable text segment is not met, the causal subgraph recalibration mechanism is triggered to re-retrieve and optimize the relevant causal path based on the latest market state until a consensus is reached. S8: Output the final hedging scheme, which includes hedging operation instructions that have been verified for consistency and its complete chain of decision-making basis, realizing the closed-loop collaborative output of strategy recommendation and interpretation generation under a unified model.

3. The method for hedging agricultural product price data through electronic data processing according to claim 1, characterized in that, In the initial dynamic causal graph structure, nodes represent price fluctuation-related factors, and edges represent statistically verified causal directions and strengths.

4. The method for hedging agricultural product price data through electronic data processing according to claim 1, characterized in that, Step S3 specifically includes: Based on the initial dynamic cause-effect graph structure generated in step S2, a recent market historical data segment is extracted using the sliding time window method. Incremental Granger causality strength re-estimation is performed on multi-source time series data within a sliding time window. The directional strength coefficient of each causal edge is dynamically updated using the recursive least squares method to obtain the corrected causal influence measurement matrix. Based on real-time market feedback data, a high-dimensional observation vector is constructed and input into the Bayesian variational inference module to estimate the changing trend of conditional probability distribution on each causal path and obtain the result of conditional probability distribution change. The modified causal influence measurement matrix and the conditional probability distribution change results are fused together to generate a joint confidence score. Based on this, the comprehensive weight parameters and node transition probabilities of each edge in the evolutionary dynamic causal graph are optimized to form a graph model parameter set. The parameter set of the graph model is subjected to sparsification and regularization processing to output an updated dynamic causal graph.

5. The method for hedging agricultural product price data through electronic data processing according to claim 4, characterized in that, The market feedback data includes fluctuations in spot prices of agricultural products, abnormal trading volumes of futures contracts, policy release event markers, and logistics disruption warning signals.

6. The method for hedging agricultural product price data through electronic data processing according to claim 1, characterized in that, Step S4 specifically includes: Based on the node set and weighted directed edge set in the evolutionary dynamic causal graph model, the underlying topology of the graph neural network is constructed to form a graph structured representation. Based on the graph structured representation, a dual-channel graph neural network architecture is initialized, which includes the policy generation channel and the interpretation generation channel. A synchronous attention gating mechanism is designed to compare the attention weight matrix of each inference step in the policy generation channel with the corresponding attention weight matrix in the interpretation generation channel in real time, and to constrain the difference between the attention weight matrix of each inference step in the policy generation channel and the corresponding attention weight matrix in the interpretation generation channel to not exceed a preset threshold through a consistency loss function. Based on the synchronized attention distribution, the optimal action probability distribution is calculated in the policy generation channel, and the state and action value functions are output; in the interpretation generation channel, the set of core causal paths that are significantly activated is identified, and the causal path activity score is output. The state and action value function and the causal path activity score are jointly input into the decision fusion module to construct a joint decision and explanation reasoning space.

7. The method for hedging agricultural product price data through electronic data processing according to claim 6, characterized in that, The initialization of the dual-channel graph neural network architecture is as follows: the policy generation channel uses a graph attention network to aggregate and calculate node features and extract key state embedding vectors for action decision-making; the interpretation generation channel deploys GAT encoders with the same structure in parallel, sharing underlying parameters and adjacency relationships.

8. The method for hedging agricultural product price data through electronic data processing according to claim 1, characterized in that, Step S5 specifically includes: Based on the node embedding vector sequence output by the evolutionary dynamic causal graph model, the attention mechanism is used to weight and aggregate the various influencing factors to calculate the joint representation vector of the current market state. The joint representation vector is input into a reinforcement learning framework based on deep deterministic policy gradients to perform action space mapping and generate a continuous hedging operation proposal sequence. At each decision time step, the explanation generation channel is activated synchronously, the set of causal edges that are significantly activated during the policy generation process is tracked, the key driving path is locked based on the synchronous attention gating mechanism, and the corresponding intermediate causal trajectory sequence is generated. The intermediate causal trajectory sequence is pruned and ranked in importance, and the core causal path set that contributes more to the current strategy action than a preset standard is selected using the causal strength threshold. The continuous hedging operation suggestion sequence and its corresponding core causal path set are used as the joint output result.

9. A method for hedging agricultural product price data through electronic data processing according to claim 8, characterized in that, The recommended sequence for continuous hedging operations includes the direction of futures position establishment, the allocation ratio of options portfolio, and the magnitude of spot position adjustment.

10. A method for hedging agricultural product price data through electronic data processing according to claim 1, characterized in that, The interpretable text fragment contains a complete logical chain of hedging action identifiers, triggering conditions, core driving factors, intermediate transmission paths, and ultimate impact inferences.

Images