Dynamic decision-making method and device for supply chain financial risk

Abstract

The invention discloses a dynamic decision-making method and device for supply chain financial risks, and the method comprises the steps: obtaining environment data corresponding to a supply chain from a plurality of data sources in real time, and carrying out the preprocessing of the environment data, and obtaining standardized data; performing risk prediction according to the environmental data through a preset risk prediction model to generate risk prediction data; performing risk conduction simulation through each agent according to the risk prediction data and the environmental data to obtain environmental change data; calculating a risk factor score according to the environment change data, and generating various risk scene data according to the risk factor score; updating the environment change data according to the risk scene data to obtain an environment updating state; and strategy optimization is carried out according to the environment updating state, and the coping strategy of each agent is determined, so that accurate prediction is carried out flexibly and rapidly according to the real-time environment data, and risk management is more accurate.

Description

Technical Field

[0001] The present invention relates to the field of machine learning technology, and in particular to a dynamic decision-making method and device for supply chain financial risks. Background Art

[0002] With the continuous expansion and complexity of the global supply chain, supply chain financial risk management faces unprecedented challenges. Supply chain finance not only involves traditional financial risks such as capital flow and credit assessment, but also includes risks brought about by various factors such as supply chain disruptions, external policy changes, and market fluctuations.

[0003] Traditional supply chain financial risk management methods often rely on static statistical models and rules and are unable to fully address complex and dynamic financial risks, especially when faced with sudden supply chain disruptions or external economic crises. These methods often lack sufficient forecasting and response capabilities.

[0004] To this end, existing technologies have begun to try to combine emerging technologies such as big data, artificial intelligence and machine learning to improve risk management. Despite this, the current system still faces several challenges, such as the timeliness and accuracy of data, the lack of adaptability of models, and the weak ability to cope with complex scenarios. Especially in the complex and changing supply chain environment, existing technologies are often unable to make accurate decision adjustments based on real-time data. Summary of the invention

[0005] The present invention provides a dynamic decision-making method and device for supply chain financial risks, which solves the technical problem that existing technologies are often unable to make accurate decision adjustments based on real-time data in a complex and changeable supply chain environment.

[0006] The present invention provides a dynamic decision-making method for supply chain financial risk, wherein the supply chain includes multiple behavioral entities, each of which has a corresponding intelligent entity; the method includes:

[0007] Acquire environmental data corresponding to the supply chain from multiple data sources in real time and preprocess them to obtain standardized data;

[0008] Perform risk prediction according to the environmental data using a preset risk prediction model to generate risk prediction data;

[0009] By each of the intelligent agents performing risk transmission simulation according to the risk prediction data and the environmental data, environmental change data is obtained;

[0010] Calculate risk factor scores according to the environmental change data, and generate multiple risk scenario data according to the risk factor scores;

[0011] The environmental change data is updated according to each of the risk scenario data to obtain an environmental update status;

[0012] The strategy is optimized according to the updated state of the environment to determine the response strategy of each of the intelligent agents.

[0013] Optionally, the method further comprises:

[0014] Obtain training environment data and historical environment data;

[0015] Inputting the training environment data and the historical environment data into a preset open source model to generate a training prediction result;

[0016] Calculating a prediction error according to the training prediction result and the actual prediction result corresponding to the historical environment data;

[0017] Back propagation is performed according to the prediction error to adjust the model parameters of the open source model until the prediction error is less than a preset error threshold, thereby obtaining a risk prediction model.

[0018] Optionally, the method further comprises:

[0019] After the response strategy is generated, each of the intelligent agents simulates and executes the response strategy to obtain environmental simulation data; the environmental simulation data includes multiple environmental factors;

[0020] Calculating a quantitative contribution value between each of the environmental factors and the original environmental factor associated with the environmental update state;

[0021] The contribution quantification values ​​are weighted and accumulated according to preset weights to obtain an overall risk assessment value;

[0022] According to the comparison result of the overall risk assessment value and the preset risk threshold, it is determined whether to output warning information.

[0023] Optionally, acquiring environmental data corresponding to the supply chain from multiple data sources in real time and preprocessing the data to obtain standardized data includes:

[0024] Acquire environmental data corresponding to the supply chain from multiple data sources in real time;

[0025] The environmental data is cleaned and standardized to obtain standardized data.

[0026] Optionally, the environmental data includes market change data and policy change data; and the step of performing risk prediction according to the environmental data using a preset risk prediction model to generate risk prediction data includes:

[0027] Inputting the market change data and the policy change data into a preset risk prediction model;

[0028] Determine the market reaction data at multiple time points according to the market change data through a risk prediction model;

[0029] By matching the market reaction data with the policy change data at the corresponding time node through the risk prediction model, the market volatility data is determined;

[0030] Risk prediction is performed according to the market volatility data through a risk prediction model to generate risk prediction data.

[0031] Optionally, the performing risk transmission simulation by each of the intelligent agents according to the risk prediction data and the environmental data to obtain environmental change data includes:

[0032] After preprocessing the risk prediction data and the environmental data by each of the intelligent agents, the data are input into the FinBERT large model built into each of the intelligent agents;

[0033] Perform semantic understanding of the environmental data through the FinBERT large model to extract financial attribute features;

[0034] Analyze the risk prediction data through the FinBERT big model to determine the risk transmission path;

[0035] The FinBERT big model uses the financial attribute characteristics and the risk transmission path, combined with historical decision data, to determine the decision behavior of each of the intelligent agents;

[0036] The environmental data is updated according to the decision-making behavior to obtain environmental change data.

[0037] Optionally, the calculating of risk factor scores according to the environmental change data and the generating of multiple risk scenario data according to the risk factor scores include:

[0038] Calling a preset linear regression model to input the environmental change data and calculate the risk factor score;

[0039] Determine the probability of occurrence of multiple risk scenarios according to a preset scenario generation formula combined with the risk factor scores;

[0040] The risk scenarios whose occurrence probability is greater than the preset scenario threshold are selected as risk scenario data.

[0041] Optionally, updating the environmental change data according to each of the risk scenario data to obtain an environmental update state includes:

[0042] Input each of the risk scenario data and the environmental change data into the preset FinBERT large model;

[0043] Determine an update decision behavior according to the risk scenario data through the FinBERT big model;

[0044] The FinBERT large model generates an environmental update state according to the interaction result between the update decision behavior and the environmental change data.

[0045] Optionally, performing strategy optimization according to the environment update state to determine the response strategy of each agent includes:

[0046] Initializing the initial response behavior of each of the intelligent agents;

[0047] According to the initial response behavior and the environmental update state, the PPO algorithm is combined to perform iterative strategy optimization, and the cumulative reward in each time step is calculated according to a preset reward formula;

[0048] When the accumulated reward reaches a maximum value, the initial response behavior of each agent at the current moment is determined as the response strategy.

[0049] A dynamic decision-making device for supply chain financial risk, wherein the supply chain includes multiple behavioral entities, each of which has a corresponding intelligent entity; the device includes:

[0050] A data acquisition module, used to acquire environmental data corresponding to the supply chain from multiple data sources in real time and pre-process it to obtain standardized data;

[0051] A big data calculation module is used to perform risk prediction according to the environmental data through a preset risk prediction model to generate risk prediction data;

[0052] An intelligent agent simulation module, used to perform risk transmission simulation according to the risk prediction data and the environmental data through each of the intelligent agents to obtain environmental change data;

[0053] A dynamic scenario generation module, used to calculate risk factor scores according to the environmental change data, and generate multiple risk scenario data according to the risk factor scores;

[0054] An environmental agent module is used to update the environmental change data according to each of the risk scenario data to obtain an environmental update status;

[0055] The strategy generation module is used to optimize the strategy according to the environment update state and determine the response strategy of each intelligent agent.

[0056] It can be seen from the above technical solutions that the present invention has the following advantages:

[0057] The present invention obtains environmental data corresponding to the supply chain from multiple data sources in real time and preprocesses it to obtain standardized data; performs risk prediction according to the environmental data through a preset risk prediction model to generate risk prediction data; performs risk transmission simulation according to the risk prediction data and environmental data by each intelligent agent to obtain environmental change data; calculates risk factor scores according to the environmental change data, and generates multiple risk scenario data according to the risk factor scores; updates the environmental change data according to each risk scenario data to obtain an environmental update status; performs strategy optimization according to the environmental update status, and determines the response strategy of each intelligent agent, thereby flexibly and quickly performing accurate predictions based on real-time environmental data, making risk management more precise. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0059] Figure 1 A flowchart of the steps of a dynamic decision-making method for supply chain financial risk provided by an embodiment of the present invention;

[0060] Figure 2 A decision block diagram of a dynamic decision-making device for supply chain financial risk provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0061] In the existing technology, many studies have attempted to improve the risk management capabilities of supply chain finance through big data analysis. For example, many big data-based models collect real-time information from all links of the supply chain (such as market prices, supplier credit, logistics conditions, etc.) to predict and evaluate risks. These methods use machine learning algorithms (such as support vector machines, decision trees, etc.) to train historical data and establish risk prediction models. However, these methods can usually only deal with relatively simple risk scenarios. For complex supply chain financial risks, such as policy changes or sudden natural disasters, it is often difficult to provide accurate predictions.

[0062] To this end, an embodiment of the present invention provides a dynamic decision-making method and device for supply chain financial risks, which is used to solve the technical problem that existing technologies are often unable to make accurate decision adjustments based on real-time data in a complex and changeable supply chain environment.

[0063] In order to make the purpose, features and advantages of the present invention more obvious and easy to understand, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described below are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0064] See also Figure 1 , Figure 1 A flowchart of the steps of a dynamic decision-making method for supply chain financial risks provided by an embodiment of the present invention.

[0065] The present invention provides a dynamic decision-making method for supply chain financial risk. The supply chain includes multiple behavioral entities, each of which has a corresponding intelligent agent. The method includes:

[0066] Step 101, acquiring environmental data corresponding to the supply chain from multiple data sources in real time and preprocessing it to obtain standardized data;

[0067] In this embodiment, multi-source data from the environment (such as market prices, transaction data, policy changes, etc.) are collected in real time, and these data provide input for risk prediction of subsequent modules.

[0068] Environmental data refers to data that measures the transaction conditions of multiple actors in the supply chain during the transaction process, including but not limited to market prices, transaction volumes or policy changes.

[0069] Among them, data sources may include but are not limited to multiple environmental data sources such as financial markets, supply chain operations, and policy changes.

[0070] In an example of the present application, step 101 may include the following sub-steps:

[0071] Obtain environmental data corresponding to the supply chain from multiple data sources in real time;

[0072] The environmental data is cleaned and standardized to obtain standardized data.

[0073] In this embodiment, after obtaining environmental data of the supply chain from multiple data sources in real time, data cleaning and standardization are performed to obtain standardized data. Specifically, multiple data collection technologies and tools, such as web crawlers, sensor access, and interface docking with various data suppliers, can be used to collect multi-source data from different environments in real time and accurately, covering real-time fluctuations in market prices, detailed records of various transaction data, and specific information on policy changes. These collected data will undergo preliminary cleaning and organization to remove invalid, duplicated, or erroneous data to ensure data quality and availability.

[0074] Step 102, performing risk prediction according to environmental data using a preset risk prediction model to generate risk prediction data;

[0075] In an example of the present application, the environmental data includes market change data and policy change data; step 102 may include the following sub-steps:

[0076] Input market change data and policy change data into the preset risk prediction model;

[0077] Determine the market reaction data at multiple time points according to market change data through risk prediction models;

[0078] Through the risk prediction model, the market reaction data is matched with the policy change data at the corresponding time node to determine the market volatility data;

[0079] Risk prediction is carried out according to market fluctuation data through the risk prediction model to generate risk prediction data.

[0080] In the embodiment of the present application, real-time market change data is obtained by cooperating with financial data suppliers. For example, stock prices and trading volume data are obtained from stock exchanges, and industry index data are obtained from industry research institutions. The obtained data is cleaned and preprocessed to remove outliers and erroneous data to ensure the quality of the data. Real-time tracking of policy documents and related announcements issued by government departments. Through text mining and data analysis technology, key information of policy changes, such as policy release time, policy type, direction and degree of policy impact on related industries, etc., is extracted. These policy change data are sorted and classified for input into the risk prediction model. The model analyzes the market change data according to its internal algorithm and predicts the market reaction data of multiple time nodes. For example, the model can predict the overall rise and fall trend of the stock market and the performance of different industries in the next month based on the current stock price fluctuations and trading volume changes. At the same time, the model matches and comprehensively analyzes the predicted market reaction data with the policy change data of the corresponding time node. For example, when the model predicts that an industry has an expansion trend in the next quarter, and the corresponding policy change data shows that the government will introduce policies to encourage the development of the industry, the model will comprehensively consider these factors to determine the market volatility data of the industry, such as the expected reduction in the volatility of the industry market. Finally, the model makes risk predictions based on the determined market volatility data and generates risk prediction data. For example, the model can predict that the probability of a financial investment portfolio losing 10%-20% in value due to market volatility in the next year is 30%.

[0081] In one example of the present application, the method further includes:

[0082] Obtain training environment data and historical environment data;

[0083] Input training environment data and historical environment data into the preset open source model to generate training prediction results;

[0084] Calculate the prediction error based on the training prediction results and the actual prediction results corresponding to the historical environment data;

[0085] Back propagation is performed according to the prediction error to adjust the model parameters of the open source model until the prediction error is less than the preset error threshold to obtain the risk prediction model.

[0086] In this embodiment, by selecting a suitable open source model, such as a financial risk prediction model based on a deep learning framework, and according to the characteristics of private data and business needs, the architecture and parameters of the model are fine-tuned in a targeted manner. Through repeated training and verification, using methods such as cross-validation and accuracy evaluation indicators, the model is continuously optimized to enable it to accurately capture the potential laws and trends of data in complex financial scenarios, significantly improve the accuracy and reliability of the model in risk prediction, and better adapt to the variability and complexity of the actual financial market.

[0087] For example, the input is the standardized training environment data and fine-tuned historical environmental data :

[0088]

[0089] in, The risk prediction results are: For private data, are the learning parameters of the model.

[0090] By fine-tuning the model with historical environmental data, it is possible to imitate the decision-making of experts and make accurate predictions when faced with risk events that have not occurred in the past.

[0091] In the specific implementation, private high-quality data (such as market data during historical financial crises, policy change data, etc.) can also be used to fine-tune the open source model to improve the model's predictive ability, especially its accuracy in complex financial scenarios. Among them, the historical environmental data can be historical financial crisis data and historical policy change data, such as market data and policy changes during the financial crisis, to help the model understand the market reaction in similar crises, and data such as monetary policies and tax policies issued by past governments, which can help the model cope with market fluctuations caused by policy changes. By fine-tuning the open source model with these data, the model can better cope with specific financial environments and emergencies. The open source model can be the LLaMA-7B large model.

[0092] Among them, LLaMA-7B is an open source large language model with 7 billion parameters released by Meta. It is designed for text generation, understanding and reasoning tasks, and is particularly suitable for environments with limited computing resources.

[0093] Step 103, each intelligent agent performs risk transmission simulation according to the risk prediction data and environmental data to obtain environmental change data;

[0094] In the embodiment of the present application, the acting subject may be a bank, enterprise or other institution. After obtaining the risk prediction data and environmental data, each intelligent agent performs risk transmission simulation according to the risk prediction data and environmental data to make corresponding decisions and interact with other intelligent agents to form a risk transmission network.

[0095] Among them, the decision formula of the agent can be:

[0096]

[0097] in, For intelligent agents behavior, For environmental data, For intelligent agents The strategy parameters, It is the expert decision data for the agent to learn, which takes into account the risk prediction data.

[0098] In addition, in the behavior training of the intelligent agent, the intelligent agent's historical decision data can be used, such as past excellent decisions: for example, which banks or companies took the right decisions during the financial crisis to avoid the spread of risks. Best practices: the most successful decision-making models in history in different market environments. These historical decision data are used to fine-tune the intelligent agent model so that it can make more practical decisions.

[0099] In an example of the present application, step 103 may include the following sub-steps:

[0100] After preprocessing the risk prediction data and environmental data through each intelligent agent, they are input into the FinBERT large model built into each intelligent agent;

[0101] Use the FinBERT large model to understand the semantics of environmental data and extract financial attribute features;

[0102] Analyze risk prediction data through the FinBERT big model to determine the risk transmission path;

[0103] The FinBERT model uses financial attribute characteristics and risk transmission paths, combined with historical decision data, to determine the decision behavior of each intelligent agent;

[0104] The environmental data is updated according to the decision-making behavior to obtain environmental change data.

[0105] In this embodiment, each agent obtains risk prediction data and environmental data respectively. Risk prediction data may include data generated by professional risk prediction models on market fluctuations, the probability of risk occurrence under policy influence, the range of losses that may be caused by risks, etc. Environmental data includes data reflecting the current state of the financial market, such as current market prices, transaction volumes, industry development trends, and macroeconomic indicators.

[0106] Each agent preprocesses the acquired risk prediction data and environmental data. For risk prediction data, this may include standardization of data formats, such as converting risk probability data from different sources into a specific probability representation format. For environmental data, this may involve data cleaning to remove outliers, such as removing unreasonable extremely high or extremely low price data caused by trading errors in market price data. At the same time, fill in missing values ​​in environmental data, such as using the mean, median, or time series-based forecast value to fill in missing parts in transaction volume data.

[0107] After preprocessing, each agent inputs the risk prediction data and environmental data into its built-in FinBERT model. The FinBERT model is a language model pre-trained specifically for the financial field. It is trained on a large amount of financial text data and can well understand the terms, concepts and semantic relationships in the financial field. Inside the FinBERT model, the input environmental data is first semantically understood. For example, when the environmental data contains the information that "the central bank conducted a reverse repurchase operation and released a certain amount of liquidity", FinBERT can accurately understand the meaning of this operation, including the direction and approximate extent of the impact on the money supply. Based on semantic understanding, the FinBERT model further extracts financial attribute features. From the above example of the central bank's reverse repurchase operation, the model may extract financial attribute features such as "increased money supply" and "abundant short-term market funds". These features can quantitatively reflect the impact of current environmental data on the financial market. The FinBERT model analyzes the input risk prediction data. Assuming that the risk prediction data indicates that the risk of corporate default in a certain industry has increased, the FinBERT model will analyze the possible transmission path of this default risk in combination with its learning and understanding of financial knowledge. For example, the model may determine that the default risk of enterprises in the industry will be transmitted through the supply chain, first affecting the collection of accounts receivable of its upstream suppliers, and then affecting the liquidity of suppliers' funds, and may even spread further to the entire industrial chain. The FinBERT big model uses the financial attribute characteristics extracted from the environmental data and the risk transmission path analyzed, and combines the historical decision data stored by each agent to determine the decision behavior of each agent. For example, for a bank agent, if the financial attribute characteristics in the environmental data show that the market is well funded and there are successful cases of expanding credit business under similar circumstances in the historical decision data, and the risk transmission path analysis shows that the current risk is controllable, then the FinBERT big model may determine that the decision behavior of the bank agent is to moderately increase the credit limit for high-quality customers. Finally, each agent updates the environmental data according to the decision behavior determined by the FinBERT big model. For example, after the above-mentioned bank agent increases the credit limit, it will cause the flow of funds in the market to change, thereby changing the relevant indicators in the environmental data such as the available funds of the enterprise and the market interest rate. These updated data are environmental change data. These environmental change data will serve as the data basis for the next round of agent decision-making, forming a dynamic decision-making cycle process.

[0108] Step 104, calculating risk factor scores according to the environmental change data, and generating multiple risk scenario data according to the risk factor scores;

[0109] In this embodiment, by simulating the environmental change data of financial markets and policies, a variety of possible risk scenario data are generated. Environmental data can be further updated to reflect environmental changes under different risk scenarios.

[0110] In an example of the present application, step 104 may include the following sub-steps:

[0111] Call the preset linear regression model to input environmental change data and calculate the risk factor score;

[0112] Determine the probability of occurrence of multiple risk scenarios based on the preset scenario generation formula combined with the risk factor scores;

[0113] Risk scenarios with a probability of occurrence greater than the preset scenario threshold are selected as risk scenario data.

[0114] The scenario generation formula can be:

[0115]

[0116] in, Indicates the situation The risk factor score of is the probability of occurrence of the scenario, n is the total number of risk scenarios, Indicates the situation risk factor scores.

[0117] In this embodiment, a linear regression model can be established, with the risk factor score as the dependent variable and the market factor and policy factor as the independent variable. The model is fitted with historical data to obtain a calculation formula for the risk factor score. It is also possible to use a machine learning algorithm such as a decision tree or a neural network, input market data and policy data, and train the model to predict the risk factor score.

[0118] Step 105, updating the environmental change data according to each risk scenario data to obtain an environmental update status;

[0119] In this embodiment, after the risk scenario data and the environmental change data are obtained, the environmental change data is updated using factors such as market fluctuations and policy changes in each risk scenario data to obtain an environmental update status.

[0120] By perceiving feedback from agent decisions and user input, adjusting market simulations, and providing necessary scenario data for other modules, it is able to perceive and respond to market changes in real time and support the optimization of subsequent strategies.

[0121] The formula is:

[0122]

[0123] in, Update the state for the environment, is the current environment data, The decision-making behavior of the intelligent agent.

[0124] In a certain risk scenario, the market fluctuates violently and there are major policy adjustments (such as a sharp rise in interest rates in a high-risk scenario). When the probability of this risk scenario reaches a certain threshold, when updating the environmental state in combination with the current environmental data and the agent's decision, this potential risk shock will be considered more carefully.

[0125] For example, if the risk scenario indicates that an industry has a high probability of facing policy tightening (high) in the future, and the current environmental data shows that the industry is in an expansion phase, and there is a large amount of investment behavior in the industry in the agent's decision-making, it may be based on this risk scenario. When updating the environmental status, the market indicators related to the industry (such as industry valuation, development expectations, etc.) are adjusted to reflect potential risk factors.

[0126] In an example of the present application, step 105 may include the following sub-steps:

[0127] Input each risk scenario data and environmental change data into the preset FinBERT big model;

[0128] Use the FinBERT big model to determine and update decision-making behaviors based on risk scenario data;

[0129] The FinBERT large model generates the environmental update status based on the interaction results between the update decision behavior and the environmental change data.

[0130] In this embodiment, after receiving the risk scenario data, each risk scenario data and environmental change data are input into the preset FinBERT big model, and the FinBERT big model performs deep semantic analysis on the input risk scenario data. For example, when the model receives a certain risk scenario data, such as "When the government introduces a new tax policy that causes the cost of a certain industry to increase by 20%, the stock price of enterprises in the industry is expected to fall by an average of 30%", the model will analyze the initial decision direction such as reducing holdings or hedging operations for investors holding stocks in the industry based on its internally stored financial knowledge and trained semantic understanding ability. The FinBERT big model further considers the interactive results of the update decision behavior and the environmental change data. For example, when an update decision behavior determined by the model is to reduce holdings of stocks in a certain industry, and the environmental change data shows that the current trading volume of the stocks in the industry has increased significantly and the price is on a downward trend, the model will analyze the impact of this interaction on the overall market environment. Based on the analysis of the above interaction results, the FinBERT big model uses its internal algorithms and logic to generate an environmental update state. This environmental update state reflects the changes that may occur in the financial market environment after considering the risk scenario and the corresponding decision behavior. For example, the environmental update status may show that after the implementation of the reduction decision, the volatility of the entire stock market will increase by 5%, the market confidence index of a specific industry will decrease by 10%, etc.

[0131] Among them, FinBERT is an open source financial optimization model based on BERT, which is specially trained to analyze financial texts (such as financial reports, market news, etc.) and is widely used in financial risk assessment and sentiment analysis tasks. The pre-trained FinBERT large model is selected. The model is trained on a large amount of financial text data and has the ability to understand professional terms, concepts and semantic relationships in the financial field.

[0132] Step 106, optimize the strategy according to the updated state of the environment and determine the response strategy of each agent.

[0133] In one example of the present application, step 106 may include the following sub-steps:

[0134] Initialize the initial response behavior of each agent;

[0135] According to the initial response behavior and the updated state of the environment, the PPO algorithm is combined to perform iterative strategy optimization, and the cumulative reward in each time step is calculated according to the preset reward formula;

[0136] When the cumulative reward reaches the maximum value, the initial response behavior of each agent at the current moment is determined as the response strategy.

[0137] Among them, the Proximal Policy Optimization (PPO) algorithm is a stable and efficient reinforcement learning algorithm, which is widely used in policy optimization in complex dynamic environments. It is particularly suitable for dealing with continuous action spaces and high-dimensional decision-making tasks in policy generation and optimization modules.

[0138] The return formula is:

[0139]

[0140] in, is the immediate reward at time step t, is the discount factor.

[0141] In an embodiment of the present application, each agent interacts with the environment according to its initial response behavior, and the state of the environment is represented by the environment update state. During the interaction process, the agent executes its initial response behavior and observes the feedback of the environment, including changes in market prices, changes in corporate operating indicators, etc., while recording the state at each time step, the behavior taken, and the immediate reward obtained. For example, in a financial investment scenario, if the initial behavior of the agent is to buy a certain stock, and in the next time step, the stock price rises, then the agent will receive a positive immediate reward, otherwise it will receive a negative reward. Set the relevant parameters of the PPO algorithm, including the learning rate, discount factor, and clipping parameters for proximal strategy optimization. The values ​​of these parameters will affect the convergence speed and optimization effect of the algorithm, and need to be tuned according to specific problems.

[0142] By initializing the agent's policy network, the input is the environment update state, and the output is the probability distribution of the agent taking different actions under the environment update state. In each iteration, the policy network is updated using the objective function of the PPO algorithm based on the collected state-behavior-reward data. The cumulative reward of each agent in each time step is calculated according to the preset reward formula. When the cumulative reward of an agent reaches the maximum value, the response behavior of each agent at this time is recorded. The initial response behavior of each agent at the current moment is determined as the final response strategy.

[0143] In one example of the present application, the method further includes:

[0144] After the response strategy is generated, each intelligent agent simulates and executes the response strategy to obtain environmental simulation data; the environmental simulation data includes a variety of environmental factors;

[0145] Calculate the quantitative value of contribution between each environmental factor and the original environmental factor associated with the environmental update state;

[0146] The contribution quantification values ​​are weighted and accumulated according to the preset weights to obtain the overall risk assessment value;

[0147] Based on the comparison results of the overall risk assessment value and the preset risk threshold, it is determined whether to output warning information.

[0148] In the embodiment of the present application, after the generation of the response strategy is completed, the effects of different strategies can also be evaluated and future risks can be predicted. Through real-time monitoring and feedback on the effectiveness of the strategy, potential risks in the market are evaluated and real-time warnings are provided. The evaluation formula is as follows:

[0149]

[0150] in, For each risk factor, The contribution of different factors (such as market fluctuations, policy changes, etc.) to the overall risk.

[0151] In this embodiment, a linear relationship model between environmental factors and the change in environmental update state can be established through linear regression analysis, and the coefficient of the model can be used as a measure of the contribution quantification value; or gray correlation analysis can be used to calculate the degree of correlation between environmental factors and key indicators in the environmental update state as the contribution quantification value. The weights are pre-set according to the importance of each environmental factor to the overall risk. These weights can be determined by expert experience, historical data statistical analysis, or machine learning algorithms. For example, in financial market risk assessment, a higher weight may be given to the key environmental factor of stock market fluctuations, such as; while a lower weight may be given to the relatively minor factor of exchange rate fluctuations.

[0152] According to the comparison result of the overall risk assessment value and the preset risk threshold, it is determined whether to output the warning information. If it is greater than or equal to the preset risk threshold, the current response strategy is judged to be likely to have a high risk, and warning information is output. Warning information can include risk warning reports, risk level identification, specific risk point analysis, etc., so that decision makers can understand the risk situation in a timely manner and take corresponding measures. If it is less than the preset risk threshold, the current response strategy can be maintained.

[0153] In addition, a user interface can be provided, which displays key indicators (such as risk scores, scenario forecasts and strategy recommendations) in intuitive graphics. Decision support includes: risk warning charts and optimization effects of response strategies. Through real-time interaction, users can directly adjust input parameters and observe system feedback to optimize supply chain finance decisions.

[0154] Compared with the prior art, the present invention has obvious advantages in many aspects. First, traditional supply chain financial risk management mostly relies on static models or decision rules based on experience, lacking flexibility and dynamic adaptability. However, the present invention introduces a large model calculation module and uses deep learning technology to process massive data in real time, which can identify potential risk patterns from complex data and make accurate predictions, making risk management more accurate and forward-looking. Secondly, the present invention simulates the interactive behaviors of various entities (such as enterprises, banks, etc.) in the supply chain through the intelligent agent simulation module, and adds three functions: information reception, decision-making, and interaction between intelligent agents, so that the system can simulate complex market behaviors and risk transmission processes, thereby improving the dynamics and complexity of risk management. In addition, the independence of the environmental agent module enables the system to perceive changes in the external environment in real time and make corresponding adjustments, enhancing the system's ability to cope with complex market environments. These innovations solve the problems of insufficient flexibility and untimely response in risk assessment and decision support in the prior art.

[0155] In an embodiment of the present application, environmental data corresponding to the supply chain is acquired in real time from multiple data sources and preprocessed to obtain standardized data; risk prediction is performed according to the environmental data through a preset risk prediction model to generate risk prediction data; risk transmission simulation is performed by each intelligent agent according to the risk prediction data and the environmental data to obtain environmental change data; risk factor scores are calculated according to the environmental change data, and multiple risk scenario data are generated according to the risk factor scores; the environmental change data is updated according to each risk scenario data to obtain an environmental update status; strategy optimization is performed according to the environmental update status to determine the response strategy of each intelligent agent, so as to flexibly and quickly make accurate predictions based on real-time environmental data, making risk management more precise.

[0156] See also Figure 2 , Figure 2 A structural block diagram of a dynamic decision-making device for supply chain financial risk provided by an embodiment of the present invention.

[0157] A dynamic decision-making device for supply chain financial risk, wherein the supply chain includes multiple behavioral entities, each of which has a corresponding intelligent entity; the device includes:

[0158] The data acquisition module 201 is used to acquire environmental data corresponding to the supply chain from multiple data sources in real time and pre-process it to obtain standardized data;

[0159] The big data calculation module 202 is used to perform risk prediction according to environmental data through a preset risk prediction model to generate risk prediction data;

[0160] The agent simulation module 203 is used to perform risk transmission simulation according to the risk prediction data and environmental data through each agent to obtain environmental change data;

[0161] A dynamic scenario generation module 204 is used to calculate risk factor scores according to environmental change data and generate multiple risk scenario data according to the risk factor scores;

[0162] The environment agent module 205 is used to update the environment change data according to each risk scenario data to obtain the environment update status;

[0163] The strategy generation module 206 is used to optimize the strategy according to the environment update status and determine the response strategy of each intelligent agent.

[0164] Optionally, the device further includes a model training module, and the model training module is used to:

[0165] Obtain training environment data and historical environment data;

[0166] Input training environment data and historical environment data into the preset open source model to generate training prediction results;

[0167] Calculate the prediction error based on the training prediction results and the actual prediction results corresponding to the historical environment data;

[0168] Back propagation is performed according to the prediction error to adjust the model parameters of the open source model until the prediction error is less than the preset error threshold to obtain the risk prediction model.

[0169] Optionally, the device further includes a risk assessment module, which is specifically used to:

[0170] After the response strategy is generated, each intelligent agent simulates and executes the response strategy to obtain environmental simulation data; the environmental simulation data includes a variety of environmental factors;

[0171] Calculate the quantitative value of contribution between each environmental factor and the original environmental factor associated with the environmental update state;

[0172] The contribution quantification values ​​are weighted and accumulated according to the preset weights to obtain the overall risk assessment value;

[0173] Based on the comparison results of the overall risk assessment value and the preset risk threshold, it is determined whether to output warning information.

[0174] Optionally, the data acquisition module 201 is specifically used for:

[0175] Obtain environmental data corresponding to the supply chain from multiple data sources in real time;

[0176] The environmental data is cleaned and standardized to obtain standardized data.

[0177] Optionally, the environmental data includes market change data and policy change data; the big data calculation module 202 is specifically used for:

[0178] Input market change data and policy change data into the preset risk prediction model;

[0179] Determine the market reaction data at multiple time points according to market change data through risk prediction models;

[0180] Through the risk prediction model, the market reaction data is matched with the policy change data at the corresponding time node to determine the market volatility data;

[0181] Risk prediction is carried out according to market fluctuation data through the risk prediction model to generate risk prediction data.

[0182] Optionally, the agent simulation module 203 is specifically used for:

[0183] After preprocessing the risk prediction data and environmental data through each intelligent agent, they are input into the FinBERT large model built into each intelligent agent;

[0184] Use the FinBERT model to understand the semantics of environmental data and extract financial attribute features;

[0185] Analyze risk prediction data through the FinBERT big model to determine the risk transmission path;

[0186] The FinBERT model uses financial attribute characteristics and risk transmission paths, combined with historical decision data, to determine the decision behavior of each intelligent agent;

[0187] The environmental data is updated according to the decision-making behavior to obtain environmental change data.

[0188] Optionally, the dynamic scenario generating module 204 is specifically used for:

[0189] Call the preset linear regression model to input environmental change data and calculate the risk factor score;

[0190] Determine the probability of occurrence of multiple risk scenarios based on the preset scenario generation formula combined with the risk factor scores;

[0191] Risk scenarios with a probability of occurrence greater than the preset scenario threshold are selected as risk scenario data.

[0192] Optionally, the environment agent module 205 is specifically used for:

[0193] Input each risk scenario data and environmental change data into the preset FinBERT big model;

[0194] Use the FinBERT big model to determine and update decision-making behaviors based on risk scenario data;

[0195] The FinBERT large model generates the environmental update status based on the interaction results between the update decision behavior and the environmental change data.

[0196] Optionally, the policy generation module 206 is specifically used for:

[0197] Initialize the initial response behavior of each agent;

[0198] According to the initial response behavior and the updated state of the environment, the PPO algorithm is combined to perform iterative strategy optimization, and the cumulative reward in each time step is calculated according to the preset reward formula;

[0199] When the cumulative reward reaches the maximum value, the initial response behavior of each agent at the current moment is determined as the response strategy.

[0200] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described devices and modules can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0201] In the several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the modules is only a logical function division. There may be other division methods in actual implementation, such as multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or modules, which can be electrical, mechanical or other forms.

[0202] The modules described as separate components may or may not be physically separated, and the components shown as modules may or may not be physical modules, that is, they may be located in one place or distributed on multiple network modules. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0203] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A dynamic decision-making method for supply chain financial risks, characterized in that: The supply chain includes a plurality of actors, each of which has a corresponding intelligent agent; the method includes: Acquire environmental data corresponding to the supply chain from multiple data sources in real time and preprocess them to obtain standardized data; Perform risk prediction according to the environmental data using a preset risk prediction model to generate risk prediction data; By each of the intelligent agents performing risk transmission simulation according to the risk prediction data and the environmental data, environmental change data is obtained; Calculate risk factor scores according to the environmental change data, and generate multiple risk scenario data according to the risk factor scores; The environmental change data is updated according to each of the risk scenario data to obtain an environmental update status; The strategy is optimized according to the updated state of the environment to determine the response strategy of each of the intelligent agents.

2. The method according to claim 1, characterized in that The method further comprises: Obtain training environment data and historical environment data; Inputting the training environment data and the historical environment data into a preset open source model to generate a training prediction result; Calculating a prediction error according to the training prediction result and the actual prediction result corresponding to the historical environment data; Back propagation is performed according to the prediction error to adjust the model parameters of the open source model until the prediction error is less than a preset error threshold, thereby obtaining a risk prediction model.

3. The method according to claim 1, characterized in that The method further comprises: After the response strategy is generated, each of the intelligent agents simulates and executes the response strategy to obtain environmental simulation data; the environmental simulation data includes multiple environmental factors; Calculating a quantitative contribution value between each of the environmental factors and the original environmental factor associated with the environmental update state; The contribution quantification values ​​are weighted and accumulated according to preset weights to obtain an overall risk assessment value; According to the comparison result of the overall risk assessment value and the preset risk threshold, it is determined whether to output warning information.

4. The method according to claim 1, characterized in that: The environmental data corresponding to the supply chain is acquired in real time from multiple data sources and preprocessed to obtain standardized data, including: Acquire environmental data corresponding to the supply chain from multiple data sources in real time; The environmental data is cleaned and standardized to obtain standardized data.

5. The method according to claim 1, characterized in that: The environmental data includes market change data and policy change data; the risk prediction is performed according to the environmental data using a preset risk prediction model to generate risk prediction data, including: Inputting the market change data and the policy change data into a preset risk prediction model; Determine the market reaction data at multiple time points according to the market change data through a risk prediction model; By matching the market reaction data with the policy change data at the corresponding time node through the risk prediction model, the market volatility data is determined; Risk prediction is performed according to the market volatility data through a risk prediction model to generate risk prediction data.

6. The method according to claim 1, characterized in that The step of performing risk transmission simulation by each of the intelligent agents according to the risk prediction data and the environmental data to obtain environmental change data includes: After preprocessing the risk prediction data and the environmental data by each of the intelligent agents, the data are input into the FinBERT large model built into each of the intelligent agents; Perform semantic understanding of the environmental data through the FinBERT large model to extract financial attribute features; Analyze the risk prediction data through the FinBERT big model to determine the risk transmission path; The FinBERT big model uses the financial attribute characteristics and the risk transmission path, combined with historical decision data, to determine the decision behavior of each of the intelligent agents; The environmental data is updated according to the decision-making behavior to obtain environmental change data.

7. The method according to claim 1, characterized in that The step of calculating the risk factor score according to the environmental change data and generating a plurality of risk scenario data according to the risk factor score includes: Calling a preset linear regression model to input the environmental change data and calculate the risk factor score; Determine the probability of occurrence of multiple risk scenarios according to a preset scenario generation formula combined with the risk factor scores; The risk scenarios whose occurrence probability is greater than the preset scenario threshold are selected as risk scenario data.

8. The method according to claim 1, characterized in that The updating of the environmental change data according to each of the risk scenario data to obtain an environmental update state includes: Input each of the risk scenario data and the environmental change data into the preset FinBERT large model; Determine an update decision behavior according to the risk scenario data through the FinBERT big model; The FinBERT large model generates an environmental update state according to the interaction result between the update decision behavior and the environmental change data.

9. The method according to claim 1, characterized in that: The step of optimizing the strategy according to the environment update state and determining the response strategy of each intelligent agent includes: Initializing the initial response behavior of each of the intelligent agents; According to the initial response behavior and the environment update state, the PPO algorithm is combined to perform iterative strategy optimization, and the cumulative reward in each time step is calculated according to a preset reward formula; When the accumulated reward reaches a maximum value, the initial response behavior of each agent at the current moment is determined as the response strategy.

10. A dynamic decision-making device for supply chain financial risks, characterized in that: The supply chain includes a plurality of behavioral entities, each of which has a corresponding intelligent entity; the device includes: A data acquisition module, used to acquire environmental data corresponding to the supply chain from multiple data sources in real time and pre-process it to obtain standardized data; A big data calculation module is used to perform risk prediction according to the environmental data through a preset risk prediction model to generate risk prediction data; An intelligent agent simulation module, used to perform risk transmission simulation according to the risk prediction data and the environmental data through each of the intelligent agents to obtain environmental change data; A dynamic scenario generation module, used to calculate risk factor scores according to the environmental change data, and generate multiple risk scenario data according to the risk factor scores; An environmental agent module is used to update the environmental change data according to each of the risk scenario data to obtain an environmental update status; The strategy generation module is used to optimize the strategy according to the environment update state and determine the response strategy of each intelligent agent.

Images