Asset transaction risk monitoring system based on data analysis

Abstract

The invention relates to the technical field of asset transaction risk monitoring, and proposes an asset transaction risk monitoring system based on data analysis, and the system comprises a data obtaining module which is used for obtaining and integrating multi-source heterogeneous data, including but not limited to transaction records, market quotation data, user behavior data, legal contract texts and financial indexes, the data is preprocessed; and the risk assessment module is connected with the data acquisition module through a data transmission technology, and assesses the risk in real time based on a preset rule and a machine learning model. Through timely acquisition and analysis of latest transaction information, market dynamics and user behavior changes, a reinforcement learning technology is utilized, and according to a real-time risk assessment result and a disposal effect, risk monitoring is ensured to be always matched with a market actual condition, so that a financial institution can timely capture a risk signal and take precautionary measures in advance, and the risk monitoring efficiency is improved. And the risk loss is effectively reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of asset trading risk monitoring, and particularly to an asset trading risk monitoring system based on data analysis. Background Art

[0002] With the continuous development and innovation of the financial market, asset trading activities have become increasingly complex and diverse, and the trading scale has been continuously expanding. This has posed unprecedented challenges to asset trading risk monitoring. For example, how to accurately determine whether a user's quota adjustment is based on the need for normal transactions or risk operations. To solve this problem, the publication number CN118552316B discloses an asset trading risk monitoring system based on data analysis, which comprehensively cooperates with a trading risk analysis module, an asset trading warning module, and a trading risk control module to predict the asset trading risks of each user and provide reaction time for timely taking control measures. However, relying only on a small amount of data such as transaction records cannot comprehensively reflect various risk factors in the asset trading process. On the one hand, the drastic fluctuations in the market conditions, abnormal user behavior patterns, dynamic changes in legal compliance requirements, and potential risks in the enterprise's financial status are difficult to effectively monitor through a single data source. On the other hand, these preset rules are often formulated based on historical experience and cannot keep up with the rapid changes in the market environment in a timely manner. When new trading patterns, policy and regulatory adjustments, or major sudden events occur in the market, traditional risk assessment models cannot quickly respond, easily leading to misjudgment or missed judgment of risks. Summary of the Invention

[0003] Aiming at the problems existing in the prior art, the purpose of the present invention is to provide an asset trading risk monitoring system based on data analysis to solve the problems raised in the above background art.

[0004] To achieve the above purpose, the present invention provides an asset trading risk monitoring system based on data analysis, including:

[0005] A data acquisition module, configured to acquire and integrate multi-source heterogeneous data, including but not limited to transaction records, market condition data, user behavior data, legal contract texts, and financial indicators, and preprocess the data;

[0006] A risk assessment module, connected to the data acquisition module through data transmission technology, and configured to perform real-time risk assessment based on preset rules and machine learning models. The risk assessment module includes:

[0007] An operation risk identification unit, configured to identify user's illegal operations through an abnormal user behavior algorithm;

[0008] A legal risk matching unit that uses natural language processing technology to analyze contract terms, match legal compliance requirements, and generate risk tags;

[0009] A market risk prediction unit that combines interest rate and exchange rate fluctuation prediction models and historical stress test data to calculate the potential loss rate of the asset portfolio;

[0010] A financial risk monitoring unit that tracks cash flow, debt ratio, and credit score in real time and simulates and evaluates the user's default probability;

[0011] An early warning and disposal module that is connected to the risk assessment module through data transmission technology and triggers a hierarchical response mechanism according to the risk assessment results, including but not limited to transaction blocking, account freezing, and pushing of manual review instructions;

[0012] A result display module that is connected to the early warning and disposal module through data transmission technology and is used to display risk results, early warning information, and disposal suggestions to users in the form of a three-dimensional scene;

[0013] An optimization processing module that is connected to the early warning and disposal module and the risk assessment module through data transmission technology and is used to dynamically optimize the preset rules and model parameters of risk analysis.

[0014] Preferably, the integration and preprocessing of multi-source heterogeneous data by the data acquisition module includes the following steps:

[0015] S11. Obtain multi-source heterogeneous data, associate data from different sources according to the relevance between the data, and merge the associated data to form a unified data set;

[0016] S12. Fill in the missing values in the data and process the outliers in the data. The formula is: , where is the mean of the data set, is the number of data points in the data set, is the th data point in the data set, is the value of the data point, is a single data point, is the standard deviation of the data set;

[0017] S13. Extract features from the processed data and perform encoding conversion to numerical data. The formula is: , where is the value of the discrete Fourier transform result at the th frequency point, is the frequency index value, is the length of the discrete sequence, is the value of the discrete-time sequence at the th sample point, is the imaginary unit, and .

[0018] Preferably, the operation risk identification unit identifies the user operation risk including the following steps:

[0019] S211. Construct an abnormal traffic pattern recognition model using a trained deep learning model , and set its initial parameters as ;

[0020] S212. Transfer some parameters in the original model to the target model. The formula is: , where are the parameters obtained by fusing the target model after migration, is the migration weight, and its value range is [0 - 1], are the parameters of the original model;

[0021] S213. Divide the feature vector dataset of operation risk into a training set , a validation set and a test set , and label each feature vector with a label indicating whether it is an abnormal traffic pattern . Use the training set and the corresponding label to train the migrated target model. The formula is: , where is the loss function value, is the number of training samples, is the true label, is the probability predicted by the model. During the training process, use the validation set to verify the model, optimize the model training by adjusting hyperparameters, and stop training when the loss function value on the validation set no longer decreases;

[0022] S214. Use the trained target model to predict the test set to obtain the predicted probability , and set a threshold. When the predicted probability is greater than the threshold, it is determined that the traffic pattern is abnormal; otherwise, it is determined to be normal;

[0023] S215. Use the test set and the corresponding predicted probability to calculate the evaluation metrics of the model and update the model regularly.

[0024] Preferably, the legal risk matching unit matches legal compliance requirements and generates risk tags, including the following steps:

[0025] S221. Match entities from contract terms and regulatory policy texts, including but not limited to legal subjects, legal events, and legal concepts, to form a collection of entities, determine the relationships between entities, and form a collection of relationships. The formula is: , where is the collection of entities, is a single entity, is the collection of relationships, is a single relationship, and build a knowledge graph based on entities and relationships to store and manage the knowledge graph. The formula is: , where is the knowledge graph constructed by entities and relationships;

[0026] S222. Convert contract terms and regulatory policy texts into vector representations and calculate the similarity between them. The formula is: , where is the contract term vector, is the regulatory policy vector, is the dot product of vectors, is the norm of the vector;

[0027] S223. According to the calculation results, determine whether the contract terms comply with the regulatory policy, and generate risk tags for the contract terms according to the judgment results;

[0028] S224. Store the contract terms with generated risk tags in the form of blockchain data blocks. Each data block includes content, timestamp, and the hash value of the previous data. The formula for calculating the hash value of the data block is: , where is the hash value of the th data block, is the hash function, is the data content of the th data block, is the hash value of the previous data block, is the timestamp.

[0029] Preferably, the market risk prediction unit calculates the potential loss rate of the asset portfolio, including the following steps:

[0030] S231. Calculate the market risk premium factor, market value factor, and book-to-market value factor. The formula is: , where , , are the market risk premium factor, market value factor, and book-to-market value factor respectively, , , , , and are the market portfolio return, risk-free return, small-cap portfolio return, large-cap portfolio return, high book-to-market ratio, and low book-to-market ratio, respectively;

[0031] S232. Perform a regression analysis on each asset to estimate the excess return coefficient of the three-factor model parameters and the sensitivity coefficients to the market risk, size factor, and value factor. The formula is: , where and are the return of the asset in the th period and the risk-free return, respectively, , , and are the excess return coefficient of the asset and the sensitivity coefficients to the market risk, size factor, and value factor, respectively, , and are the returns of the market risk premium factor, market value factor, and book-to-market factor in the th period, respectively, is the error term;

[0032] S233. Assist in asset correlation analysis and decision-making by aggregating neighbor asset features and generating asset embedding vectors. The propagation formula is: , where and are the asset node feature matrices of the th layer and the th layer, respectively, is the activation function, is the degree matrix, is the adjacency matrix after adding self-loops, is the scientific department weight matrix of the th layer. And train the propagation formula by minimizing the loss function between the predicted value and the true value. The formula is: , where is the mean squared error loss value, is the true risk value of asset , is the risk value predicted by the propagation formula;

[0033] S234. Calculate the correlation score between assets using the trained propagation formula. The formula is: , where is the correlation score between asset and The Euclidean distance between node embedding vectors and are the values of the and th dimensions of the node embedding vectors and respectively, where is the dimension of the embedding vector. The Monte Carlo simulation method is adopted. The initial risky assets are set, and the next affected asset is randomly selected according to the correlation score. The simulation process is repeated multiple times, and the probability of each asset being affected is statistically calculated;

[0034] S235. Convert the asset correlation matrix and risk propagation rules into quantum states and quantum gate operations, use quantum algorithms to accelerate the calculation of risk diffusion paths, and read the quantum computing results through quantum measurement operations and convert them into classical risk diffusion path information.

[0035] Preferably, the financial risk monitoring unit simulates and evaluates the default probability of users, including the following steps:

[0036] S241. Divide the preprocessed data into institutions , and each institution has a local data set . In the formula, is the feature vector of the th sample, is the corresponding label, is the number of samples of a single institution . Initialize a credit scoring model locally for each institution, and set the model parameters to be initialized as ;

[0037] S242. Use the logarithmic loss function to train a single institution using the local data set . The formula is: . In the formula, is the prediction function of logistic regression, and , where is the transpose of the parameter of the credit scoring model of a single institution , and is the input sample feature vector;

[0038] S243. A single institution updates the parameters of the credit scoring model through the gradient descent optimization algorithm. The formula is: , where, is the learning rate, is the number of training epochs;

[0039] S244. When a single institution uploads the model parameters, add Laplace noise to the parameters. The formula is: , where, and are the original model vector parameters of a single institution and the th element after adding Laplace noise, is the noise value sampled from the Laplace distribution, is the sensitivity of the model parameters, is the privacy budget;

[0040] S245. After a single institution uploads the model parameters with added noise, perform an aggregation operation by weighted average. The formula for the aggregated model parameters is: , where, is the global model parameters obtained after rounds of aggregation operations, is the number of institutions in federated learning, and distribute the aggregated model parameters to each institution , and each institution updates the credit scoring model using the aggregated model parameters ;

[0041] S246. Use the trained global model to predict defaults for new samples and evaluate the model using the test dataset.

[0042] Preferably, the optimization processing module dynamically optimizes the risk analysis preset rules and model parameters, including the following steps:

[0043] S51. Determine the reinforcement learning elements by defining the state space, action space, and reward function;

[0044] S52. Establish a policy network of a neural network , where, are the network parameters, is the action taken, is the network input state, output the probability distribution of taking each action in this state, and randomly initialize the parameters of the policy network;

[0045] S53. Interact with the environment and collect experiences, update the parameters of the policy network and update the parameters of the value network;

[0046] S54. Repeat the loop of step S53 multiple times so that the agent can learn the optimal policy, thereby dynamically optimizing the preset rules and machine learning model parameters in the risk assessment module.

[0047] Preferably, in step S51, the state space combines the real-time risk assessment results of the system and environmental information, including but not limited to operation risk scores, legal risk scores, market risk scores, financial risk scores, market volatility, and interest rates, and a state space collection is established. The action space is the actions to adjust the risk assessment module, including but not limited to adjusting the parameters of the preset rules and updating the parameters in the machine model. The reward function is used to measure the quality of taking action in state , and the formula is: , where is the accuracy and disposal effect of risk assessment, is the improvement in risk assessment accuracy after taking action , is the improvement in disposal efficiency, and are the corresponding weight coefficients, and .

[0048] Preferably, in step S53, updating the parameters of the policy network and updating the parameters of the value network include the following steps:

[0049] S531. At each time step , based on the current state , the agent uses the policy network to sample an action greedily in combination with , and after applying it to the risk assessment module, a new state and a reward are obtained, and the experience tuple is stored in the experience replay buffer to collect experience;

[0050] S532. Randomly sample a batch of experience tuples from the experience replay buffer , and the formula is: , where is the batch size, is the combination of experience tuples;

[0051] S533. Calculate the advantage function to measure the advantage of taking action in state relative to the average situation, and the formula is: , where is a value network for estimating the state value in where

[0052] S534. Update the parameters of the policy network using the policy gradient algorithm, and use the gradient ascent method to update the parameters. The formula is: where is the state when taking the action the logarithm of the output probability of the policy network with respect to the parameter gradient, is the learning rate;

[0053] S535. Optimize the value network to provide a reliable reference for updating the policy network;

[0054] S536. Continuously repeat steps S532 - S535 to continuously update the parameters of the policy network and the value network until the agent learns the optimal policy.

[0055] Preferably, in step S535, to optimize the value network, after calculating the target value, use the gradient square weighted algorithm to adjust the learning rate and update the network parameters. The formula is: where is the target value, and are respectively and the weighted moving average of the gradient squares at times is the decay coefficient, is the gradient at time squared, is after the model parameter value at time updated the model parameter value at time is the initial learning rate, is the denominator term for adjusting the learning rate, is a constant with a value of , is the gradient of the model parameters at time

[0056] The beneficial effects of an asset trading risk monitoring system based on data analysis provided by the present invention are:

[0057] 1. Through an all-round data collection system that covers multi-source heterogeneous data such as transaction records, market conditions, user behavior, legal contract texts, and financial indicators, during the risk analysis process, by combining preset rules with machine learning models, potential risk factors behind the data are fully explored. The machine learning model can automatically learn complex data patterns and rules, and conduct real-time and accurate assessments of operational risks, legal risks, market risks, and financial risks. Compared with traditional single data sources and simple assessment models, the present invention greatly improves the accuracy of risk assessment, reduces the situations of risk misjudgment and missed judgment, and provides a more reliable basis for risk decision-making for financial institutions.

[0058] 2. By timely obtaining and analyzing the latest transaction information, market dynamics, and changes in user behavior, using reinforcement learning technology, according to real-time risk assessment results and disposal effects, the preset rules and machine learning model parameters in the risk assessment module are dynamically optimized. When the market environment changes or new risk characteristics appear, the system can quickly make adjustments, automatically update risk assessment criteria and warning thresholds, and ensure that risk monitoring always matches the actual market situation. This dynamic adaptive ability enables financial institutions to timely capture risk signals, take preventive measures in advance, and effectively reduce risk losses. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or in the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0060] Figure 1 It is a schematic diagram of the system module operation of an asset trading risk monitoring system based on data analysis provided by this application;

[0061] Figure 2 It is a schematic diagram of the risk assessment module system module of an asset trading risk monitoring system based on data analysis provided by this application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0062] The following will further describe in detail the specific embodiments of the present invention in conjunction with the drawings in the specification and the embodiments. The following embodiments are only used to illustrate the present invention, but cannot be used to limit the scope of the present invention.

[0063] As Figure 1 - Figure 2 shown, this embodiment proposes an asset trading risk monitoring system based on data analysis, including:

[0064] A data acquisition module, which is used to acquire and integrate multi-source heterogeneous data, including but not limited to transaction records, market quotation data, user behavior data, legal contract texts, and financial indicators, and preprocess the data;

[0065] A risk assessment module, which is connected to the data acquisition module through data transmission technology. Based on preset rules and machine learning models, it conducts real-time risk assessment. The risk assessment module includes:

[0066] An operational risk identification unit, which identifies users' illegal operations through user behavior anomaly algorithms;

[0067] A legal risk matching unit, which uses natural language processing technology to parse contract terms, match legal compliance requirements, and generate risk tags;

[0068] A market risk prediction unit, which combines interest rate and exchange rate fluctuation prediction models and historical stress test data to calculate the potential loss rate of the asset portfolio;

[0069] A financial risk monitoring unit, which real-time tracks cash flow, debt ratio, and credit score, and simulates and evaluates the user's default probability;

[0070] An early warning and disposal module, which is connected to the risk assessment module through data transmission technology. According to the risk assessment results, it triggers a hierarchical response mechanism, including but not limited to transaction blocking, account freezing, and pushing of manual review instructions;

[0071] A result display module, which is connected to the early warning and disposal module through data transmission technology, and is used to display risk results, early warning information, and disposal suggestions to users in the form of a three-dimensional scene;

[0072] An optimization processing module, which is connected to the early warning and disposal module and the risk assessment module through data transmission technology, and is used to dynamically optimize the preset rules and model parameters of risk analysis.

[0073] Specifically, through an all-round data acquisition system covering multi-source heterogeneous data such as transaction records, market quotations, user behavior, legal contract texts, and financial indicators, in the process of risk analysis, combining preset rules with machine learning models, fully mining the potential risk factors behind the data. The machine learning model can automatically learn complex data patterns and rules, and conduct real-time and accurate assessment of operational risk, legal risk, market risk, and financial risk. Compared with traditional single data sources and simple assessment models, the present invention greatly improves the accuracy of risk assessment, reduces the situations of risk misjudgment and missed judgment, and provides a more reliable basis for risk decision-making for financial institutions.

[0074] In this embodiment, the integration and preprocessing of multi-source heterogeneous data by the data acquisition module include the following steps:

[0075] S11. Obtain multi-source heterogeneous data, associate data from different sources according to the relevance between the data, and merge the associated data to form a unified data set;

[0076] S12. Fill in the missing values in the data and process the outliers in the data. The formula is: , where is the mean of the data set, is the number of data points in the data set, is the th data point in the data set, is the value of the data point, is a single data point, is the standard deviation of the data set;

[0077] S13. Extract features from the processed data and perform encoding conversion to numerical data. The formula is: , where is the value of the discrete Fourier transform result at the th frequency point, is the frequency index value, is the length of the discrete sequence, is the value of the discrete time series at the th sample point, is the imaginary unit, and .

[0078] In this embodiment, the operation risk identification unit identifies the user operation risk including the following steps:

[0079] S211. Construct an abnormal traffic pattern recognition model using a trained deep learning model , and set its initial parameters as ;

[0080] S212. Transfer some of the parameters in the original model to the target model. The formula is: , where are the parameters obtained by fusing the target model after migration, is the migration weight, and its value range is [0 - 1], are the parameters of the original model;

[0081] S213. Divide the feature vector data set of operation risk into a training set , a validation set and a test set , and label each feature vector with a label indicating whether it is an abnormal traffic pattern , use the training set and the corresponding labels to train the migrated target model. The formula is: , where is the loss function value, is the number of training samples, is the true label, is the probability predicted by the model. During the training process, use the validation set to validate the model. Optimize the model training by adjusting the hyperparameters. When the loss function value on the validation set no longer decreases, stop the training;

[0082] S214. Use the trained target model to predict the test set to obtain the predicted probability , and set a threshold. When the predicted probability is greater than the threshold, it is determined that the traffic pattern is abnormal; otherwise, it is determined to be normal;

[0083] S215. Use the test set and the corresponding predicted probability to calculate the evaluation metrics of the model and update the model regularly.

[0084] In this embodiment, the legal risk matching unit matches the legal compliance requirements and generates risk labels, including the following steps:

[0085] S221. Match the entities from the contract terms and regulatory policy texts, including but not limited to legal subjects, legal events, and legal concepts, to form an entity set, determine the relationships between the entities, and form a relationship set. The formula is: , where is the entity set, is a single entity, is the relationship set, is a single relationship, and build a knowledge graph based on the entities and relationships to store and manage the knowledge graph. The formula is: , where is the knowledge graph constructed from entities and relationships;

[0086] S222. Convert the contract terms and regulatory policy texts into vector representations and calculate the similarity between them. The formula is: , where is the contract term vector, is the regulatory policy vector, is the dot product of vectors, is the norm of the vector;

[0087] S223. According to the calculation results, determine whether the contract terms comply with the regulatory policies, and based on the judgment results, generate risk labels for the contract terms;

[0088] S224. Store the contract terms with generated risk labels in the form of data blocks on the blockchain. Each data block includes content, a timestamp, and the hash value of the previous data. The formula for calculating the hash value of the data block is: , where is the hash value of the -th data block, is the hash function, is the -th data block's data content, is the hash value of the previous data block, is the timestamp.

[0089] In this embodiment, the steps for the market risk prediction unit to calculate the potential loss rate of the asset portfolio include the following:

[0090] S231. Calculate the market risk premium factor, market value factor, and book-to-market factor. The formula is: , where , , are the market risk premium factor, market value factor, and book-to-market factor respectively, , , , , and are the market portfolio return, risk-free return, small market value portfolio return, large market value portfolio return, high book-to-market ratio, and low book-to-market ratio respectively;

[0091] S232. Conduct a regression analysis on each asset to estimate the excess return coefficient of the three-factor model parameters and the sensitivity coefficients of market risk, size factor, and value factor. The formula is: , where and are the return and risk-free return of the asset in the -th period respectively, , , and are the excess return coefficient of the asset and the sensitivity coefficients of market risk, size factor, and value factor respectively, , and are the returns of the market risk premium factor, market value factor, and book-to-market factor in the -th period respectively, is the error term;

[0092] S233. By aggregating neighbor asset features and generating asset embedding vectors, it assists in asset correlation analysis and decision-making. Its propagation formula is: , where in the formula, and are the asset node feature matrices of the -th layer and respectively, is the activation function, is 's degree matrix, is the adjacency matrix after adding self-loops, is the scientific department weight matrix of the -th layer, and the propagation formula is trained by minimizing the loss function between the predicted value and the true value. The formula is: , where in the formula, is the mean squared error loss value, is the true risk value of asset , is the risk value predicted by the propagation formula;

[0093] S234. Using the trained propagation formula, calculate the correlation score between assets. The formula is: , where in the formula, is the Euclidean distance between the node embedding vectors of assets and , and are the -th and -th dimension values of the node embedding vectors of assets and respectively, is the dimension of the embedding vector, and the Monte Carlo simulation method is adopted. Set the initial risky assets, randomly select the next affected asset according to the correlation score, repeat the simulation process multiple times, and count the probability of each asset being affected;

[0094] S235. Convert the asset correlation matrix and risk propagation rules into quantum states and quantum gate operations, use quantum algorithms to accelerate the calculation of the risk diffusion path, and read the quantum calculation results through quantum measurement operations and convert them into classical risk diffusion path information.

[0095] In this embodiment, the financial risk monitoring unit simulates and evaluates the default probability of users, including the following steps:

[0096] S241. Divide the preprocessed data into institutions , each institution has a local dataset , where is the feature vector of the th sample, is the corresponding label, is the number of samples of a single institution , and a credit scoring model is initialized locally for each institution , and the model parameters are initialized to ;

[0097] S242. Use the logarithmic loss function to train a single institution using the local dataset , and the formula is: , where is the prediction function of logistic regression, and , where is the transpose of the parameters of the credit scoring model of a single institution , and is the input sample feature vector;

[0098] S243. A single institution updates the parameters of the credit scoring model through the gradient descent optimization algorithm, and the formula is: , , where is the learning rate, is the number of training rounds;

[0099] S244. When uploading the model parameters of a single institution , add Laplace noise to the parameters, and the formula is: , where and are the original model vector parameters and the th element after adding Laplace noise of a single institution , is the noise value sampled from the Laplace distribution, is the sensitivity of the model parameters,

[0100] S245. A single institution After uploading the model parameters with added noise, an aggregation operation is performed by weighted averaging. The formula for the aggregated model parameters is: , In the formula, is the global model parameters obtained after rounds of aggregation operations, is the number of institutions in federated learning, and the aggregated model parameters are distributed to each institution . Each institution updates the credit scoring model with the aggregated model parameters ; ;

[0101] S246. Use the trained global model to predict defaults for new samples and evaluate the model using the test dataset.

[0102] In this embodiment, the steps for the optimization processing module to dynamically optimize the risk analysis preset rules and model parameters are as follows:

[0103] S51. Determine the reinforcement learning elements by defining the state space, action space, and reward function;

[0104] S52. Establish a policy network of a neural network , where are the network parameters, is the action taken, is the network input state, outputting the probability distribution of taking each action in this state, and randomly initializing the parameters of the policy network ;

[0105] S53. Interact with the environment and collect experiences, update the parameters of the policy network and update the parameters of the value network;

[0106] S54. Repeat step S53 multiple times so that the agent can learn the optimal policy, thereby dynamically optimizing the preset rules and machine learning model parameters in the risk assessment module.

[0107] In this embodiment, in step S51, the state space combines the real-time risk assessment results of the system and environmental information, including but not limited to operational risk scores, legal risk scores, market risk scores, financial risk scores, market volatility, and interest rates, and establishes a state space set. The action space is the actions to adjust the risk assessment module, including but not limited to adjusting the parameters of the preset rules and updating the parameters in the machine model. The reward function is used to measure the quality of taking action in state , and the formula is: , in the formula, For the accuracy of risk assessment and the effectiveness of disposal, For taking actions The improvement of the accuracy of risk assessment after For the improvement of disposal efficiency, And For the corresponding weight coefficients, and .

[0108] In this embodiment, in step S53, updating the parameters of the policy network and updating the parameters of the value network include the following steps:

[0109] S531. The agent, at each time step Based on the current state , uses the policy network Combined with The greedy policy to sample actions , and after applying it to the risk assessment module, obtains a new state And a reward , and stores the experience tuple Into the experience replay buffer To collect experience;

[0110] S532. Randomly sample a batch of experience tuples from the experience replay buffer , and the formula is: , where Is the batch size, Is the experience tuple combination;

[0111] S533. Calculate the advantage function To measure the advantage of taking an action In the state Relative to the average case, and the formula is: , where Is a value network used to estimate The state value in, Is the discount factor;

[0112] S534. Use the policy gradient algorithm to update the parameters of the policy network, and use the gradient ascent method to update the parameters, and the formula is: , where Is the state When taking an action , the logarithm of the output probability of the policy network with respect to the parameter Gradient of, Is the learning rate;

[0113] S535. Optimize the value network to provide a reliable reference for the update of the policy network;

[0114] S536. Continuously repeat steps S532 - S535 to continuously update the parameters of the policy network and the value network until the agent learns the optimal policy.

[0115] In this embodiment, in step S535, after calculating the target value for optimizing the value network, the gradient square weighted algorithm is used to adjust the learning rate and update the network parameters. The formula is: , where in the formula, is the target value, and are respectively and the weighted moving average of the gradient squares at times is the decay coefficient, is the square of the gradient at time is the model parameter value after updating at time and is the model parameter value at time is the initial learning rate, is the denominator term for adjusting the learning rate, is a constant with a value of , is the gradient of the model parameters at time

[0116] Specifically, by timely obtaining and analyzing the latest transaction information, market dynamics, and user behavior changes, using reinforcement learning technology, according to the real-time risk assessment results and disposal effects, dynamically optimize the preset rules and machine learning model parameters in the risk assessment module. When the market environment changes or new risk characteristics appear, the system can quickly make adjustments, automatically update the risk assessment criteria and warning thresholds, ensuring that the risk monitoring always matches the actual market situation. This dynamic adaptive ability enables financial institutions to timely capture risk signals, take preventive measures in advance, and effectively reduce risk losses.

[0117] The above embodiments are only used to illustrate the present invention, rather than limiting the present invention. Although the present invention has been described in detail with reference to the embodiments, those of ordinary skill in the art should understand that various combinations, modifications, or equivalent replacements of the technical solutions of the present invention do not depart from the spirit and scope of the technical solutions of the present invention, and should all be covered within the scope of the claims of the present invention.

Claims

1. An asset transaction risk monitoring system based on data analysis, characterized in that: include: Data acquisition module, used to acquire and integrate multi-source heterogeneous data, including but not limited to transaction records, market data, user behavior data, legal contract texts, and financial indicators, and pre-process the data; The risk assessment module is connected to the data acquisition module through data transmission technology, and performs real-time risk assessment based on preset rules and machine learning models. The risk assessment module includes: Operational risk identification unit, which identifies user illegal operations through user behavior anomaly algorithms; The legal risk matching unit uses natural language processing technology to parse contract terms, match legal compliance requirements and generate risk tags; The market risk prediction unit calculates the potential loss rate of the asset portfolio by combining the interest rate and exchange rate fluctuation prediction model and historical stress test data; Financial risk monitoring unit, which tracks cash flow, debt ratio and credit score in real time, and simulates and evaluates the user's probability of default; The early warning and disposal module is connected to the risk assessment module through data transmission technology, and triggers a graded response mechanism based on the risk assessment results, including but not limited to transaction blocking, account freezing, and manual review instruction push; The result display module is connected to the early warning and disposal module through data transmission technology, and is used to display risk results, early warning information and disposal suggestions to users in the form of three-dimensional scenes; The optimization processing module is connected to the early warning and disposal module and the risk assessment module through data transmission technology, and is used to dynamically optimize the preset rules and model parameters of risk analysis.

2. The asset transaction risk monitoring system based on data analysis according to claim 1 is characterized in that: The data acquisition module integrates and preprocesses multi-source heterogeneous data, including the following steps: S11. Obtain multi-source heterogeneous data, associate data from different sources according to the correlation between the data, and merge the associated data to form a unified data set; S12. Fill the missing values ​​in the data and process the outliers in the data. The formula is: ， In the formula, is the mean of the data set, is the number of data points in the dataset, For the data set data points, For data points value, For a single data point, is the standard deviation of the data set; S13, extract features from the processed data and convert them into numerical data by encoding. The formula is: ， In the formula, The discrete Fourier transform result is The value at the frequency point, is the frequency index value, is the length of the discrete sequence, For a discrete time series The value at the sample points, is an imaginary unit, and .

3. The asset transaction risk monitoring system based on data analysis according to claim 1 is characterized in that: The operation risk identification unit identifies the user operation risk including the following steps: S211. Use the trained deep learning model to build an abnormal traffic pattern recognition model , and set its initial parameters to ; S212. Migrate some parameters in the original model to the target model. The formula is: ， In the formula, The parameters obtained by fusion of the target model after migration, is the migration weight, the value range is [0-1], are the parameters of the original model; S213. Divide the feature vector data set of operational risk into a training set , validation set and test set , and label each feature vector as an abnormal traffic pattern , using the training set and the corresponding labels The formula for training the migrated target model is: ， In the formula, is the loss function value, is the number of training samples, is the true label, The probability predicted by the model and the validation set used during training Verify the model and optimize model training by adjusting hyperparameters. Stop training when the loss function value on the validation set no longer decreases. S214. Use the trained target model For the test set Make predictions and get prediction probabilities , and set a threshold when the predicted probability If it is greater than the threshold, the traffic pattern is judged to be abnormal, otherwise, it is judged to be normal; S215. Use the test set and the corresponding predicted probability Calculate the evaluation indicators of the model and update the model regularly.

4. The asset transaction risk monitoring system based on data analysis according to claim 1 is characterized in that: The legal risk matching unit matches the legal compliance requirements and generates risk tags, including the following steps: S221. Match entities from contract terms and regulatory policy texts, including but not limited to legal entities, legal events, and legal concepts, to form an entity collection, determine the relationship between entities, and form a relationship collection. The formula is: ， In the formula, For entity collection, For a single entity, For the relationship collection, For a single relationship, a knowledge graph is built based on entities and relationships to store and manage the knowledge graph. The formula is: , where Build knowledge graphs for entities and relationships; S222. Convert the contract terms and regulatory policy texts into vector representations and calculate the similarity between them. The formula is: ， In the formula, is the contract terms vector, is the regulatory policy vector, is the vector dot product, is the magnitude of the vector; S223. Determine whether the contract terms comply with regulatory policies based on the calculation results, and generate risk labels for the contract terms based on the determination results; S224. The contract terms for generating risk tags are stored in the form of data blocks of the blockchain. Each data block includes content, timestamp, and hash value of the previous data. The hash value formula for calculating the data block is: ， In the formula, For the The hash value of a data block, is a hash function, For the The data content of each data block, is the hash value of the previous data block, Is the timestamp.

5. The asset transaction risk monitoring system based on data analysis according to claim 1 is characterized in that: The market risk prediction unit calculates the potential loss rate of the asset portfolio by the following steps: S231. Calculate the market risk premium factor, market value factor and book value factor using the following formula: ， In the formula, , , They are market risk premium factor, market value factor and book value factor. , , , , and They are market portfolio return, risk-free rate of return, small market capitalization portfolio return, large market capitalization portfolio return, high book-to-market ratio and low book-to-market ratio; S232. Perform regression analysis on each asset to estimate the excess return coefficient of the three-factor model parameters and the sensitivity coefficients of market risk, scale factor, and value factor. The formula is: ， In the formula, and The assets are The return and risk-free return of the period, , , and They are the excess return coefficient of the asset and the sensitivity coefficients of market risk, scale factor and value factor, , and Respectively The returns of the market risk premium factor, market value factor and book value factor are is the error term; S233, by aggregating neighbor asset features and generating asset embedding vectors, assists asset correlation analysis and decision-making, and its propagation formula is: , where and Respectively Layer and The asset node feature matrix, is the activation function, for The degree matrix of is the adjacency matrix after adding the self-loop, For the The science system weight matrix of the layer is trained by minimizing the loss function between the predicted value and the true value. The formula is: ， In the formula, is the mean square error loss value, For assets The true risk value, is the risk value predicted by the propagation formula; S234. Use the trained propagation formula to calculate the correlation score between assets. The formula is: ， In the formula, For assets and The Euclidean distance between node embedding vectors, and Assets and Node Embedding Vector and No. The value of the dimension, is the dimension of the embedding vector, and the Monte Carlo simulation method is used to set the initial risk asset, randomly select the next affected asset according to the correlation score, repeat the simulation process multiple times, and count the probability of each asset being affected; S235. Convert the asset correlation matrix and risk propagation rules into quantum states and quantum gate operations, use quantum algorithms to accelerate the calculation of risk diffusion paths, and read quantum calculation results through quantum measurement operations to convert them into classical risk diffusion path information.

6. The asset transaction risk monitoring system based on data analysis according to claim 1 is characterized in that: The financial risk monitoring unit simulates and evaluates the default probability of the user, including the following steps: S241, the preprocessed data is divided into Institutions , each institution Have a local dataset , In the formula, For the The feature vector of the samples, For the corresponding label, For a single institution The number of samples is set, and a credit scoring model is initialized locally for each institution. , let the model parameters be initialized as ; S242, using logarithmic loss function for a single institution Using local datasets For training, the formula is: ， In the formula, is the prediction function of logistic regression, and ,in, For a single institution Credit scoring model parameter The transpose of is the input sample feature vector; S243, Single institution Updating the credit scoring model via the gradient descent optimization algorithm parameter , the formula is: ， In the formula, is the learning rate, is the number of training rounds; S244. In a single institution When uploading model parameters, add Laplace noise to the parameters. The formula is: ， In the formula, and For each institution Original model vector parameters and the first elements, is the noise value sampled from the Laplace distribution, is the sensitivity of the model parameters, Budget for privacy; S245, Single institution After uploading the model parameters with added noise, the aggregation operation is performed by weighted averaging. The formula of the model parameters after aggregation is: ， In the formula, For passing The global model parameters obtained after the round aggregation operation, is the number of institutions in federated learning, and the aggregated model parameters Distribute to various institutions , various institutions Use the aggregated model parameters Update credit scoring model ; S246. Use the trained global model to predict defaults for new samples and use the test dataset to evaluate the model.

7. The asset transaction risk monitoring system based on data analysis according to claim 1 is characterized in that: The optimization processing module dynamically optimizes the risk analysis preset rules and model parameters including the following steps: S51. Determine the elements of reinforcement learning by defining the state space, action space and reward function; S52. Build a neural network The policy network of are network parameters, For the actions taken, Input a state to the network, output the probability distribution of taking each action in that state, and randomly initialize the parameters of the policy network ; S53, interact with the environment and collect experience, update the parameters of the policy network and update the parameters of the value network; S54, repeating step S53 multiple times, so that the intelligent agent can learn the optimal strategy, thereby dynamically optimizing the preset rules and machine learning model parameters in the risk assessment module.

8. The asset transaction risk monitoring system based on data analysis according to claim 7 is characterized in that: In step S51, the state space is to combine the real-time risk assessment results of the system with environmental information, including but not limited to operational risk scores, legal risk scores, market risk scores, financial risk scores, market volatility and interest rates, and establish a state space collection. The action space is to adjust the actions of the risk assessment module, including but not limited to adjusting the parameters of the preset rules and updating the parameters in the machine model. The reward function is used to measure the state space. Take action The formula is: , In the formula, To ensure the accuracy of risk assessment and the effectiveness of treatment, To take action Improvement of post-risk assessment accuracy, To improve the processing efficiency, and is the corresponding weight coefficient, and .

9. The asset transaction risk monitoring system based on data analysis according to claim 7 is characterized in that: In step S53, updating the parameters of the policy network and updating the parameters of the value network include the following steps: S531, the agent at each time step Based on current status , using the policy network Combination Greedy strategy sampling action , applied to the risk assessment module to obtain the new status and rewards , and the experience tuple Save experience replay buffer To gather experience; S532: Replaying from the Experience Buffer A batch of experience tuples are randomly sampled in the formula: ， In the formula, is the batch size, is the combination of experience tuples; S533. Calculate advantage function Measure in state Take action The advantage over the average is given by: ， In the formula, is a value network used to estimate The state value in is the discount factor; S534, update the parameters of the policy network using the policy gradient algorithm, and update the parameters using the gradient ascent method, the formula is: ， In the formula, Status Take action When , the logarithm of the policy network output probability is about the parameter The gradient of is the learning rate; S535, optimize the value network and provide a reliable reference for updating the strategy network; S536. Repeat steps S532 to S535 continuously so that the parameters of the policy network and the value network are continuously updated until the intelligent agent learns the optimal strategy.

10. The asset transaction risk monitoring system based on data analysis according to claim 9, characterized in that: In step S535, after calculating the target value, the optimization value network uses the gradient square weighted algorithm to adjust the learning rate and update the network parameters. The formula is: ， In the formula, is the target value, and They are and The weighted moving average of the squared gradient at each moment, is the attenuation coefficient, for Time gradient The square of For passing Model parameter values ​​at time After Update The model parameter values ​​at time , is the initial learning rate, To adjust the denominator of the learning rate, is a constant, and its value is , for The gradient of the model parameters at time t.

Images