Credit risk monitoring method based on model-free adaptive control

Abstract

The invention is suitable for the technical field of risk management, and particularly relates to a credit risk monitoring method based on model-free adaptive control, and the method comprises the steps: obtaining personal credit data, carrying out the preprocessing of the personal credit data, and obtaining the preprocessing data; constructing a model-free adaptive risk monitoring framework based on the preprocessed data; obtaining to-be-processed credit data, extracting risk characteristics of the credit data in real time through a dynamic learning module, calculating an influence degree of credit data change on a risk early warning signal by using a pseudo-partial derivative calculation module, and dynamically adjusting a risk monitoring strategy based on the influence degree; constructing a risk prediction model, and predicting and evaluating the credit risk in real time; and the internal structure and parameters of the model-free adaptive risk monitoring framework are optimized through an online learning algorithm. According to the method, the robustness and the adaptive capacity of the control system are improved, the dependence on prior knowledge of the system is remarkably reduced, and the control strategy is more flexible and efficient.

Description

Technical Field

[0001] The present invention belongs to the technical field of risk management, and in particular relates to a credit risk monitoring method based on model-free adaptive control. Background Art

[0002] Model-free adaptive control (MFAC) is an advanced control method that can achieve adaptive control without building a mathematical model of the controlled system.

[0003] With the continuous development of model-free adaptive control (MFAC) technology, its application in industrial automation, financial risk management, and many complex system control fields is becoming more and more extensive. However, due to the uncertainty of system characteristics and operating environment, the control process often faces problems such as incomplete data or inaccurate models, which will affect the formulation and implementation of control strategies. At present, commonly used control methods such as model-based control and PID control are often limited in effect when facing highly nonlinear and time-varying systems, and require cumbersome model building processes. Summary of the invention

[0004] The purpose of the present invention is to provide a credit risk monitoring method based on model-free adaptive control, aiming to solve the problem that commonly used control methods such as model-based control and PID control are often limited in effect when facing highly nonlinear and time-varying systems, and require a cumbersome model building process.

[0005] The present invention is implemented as follows: a credit risk monitoring method based on model-free adaptive control, the method comprising:

[0006] Acquire personal credit data, pre-process the personal credit data, and obtain pre-processed data;

[0007] Building a model-free adaptive risk monitoring framework based on preprocessed data, wherein the model-free adaptive risk monitoring framework includes a dynamic learning module and a pseudo partial derivative calculation module;

[0008] Obtaining credit data to be processed, extracting risk features of the credit data in real time through a dynamic learning module, calculating the degree of influence of changes in credit data on risk warning signals using a pseudo partial derivative calculation module, and dynamically adjusting risk monitoring strategies based on the degree of influence, wherein the degree of influence is represented by a pseudo partial derivative;

[0009] Build risk prediction models to predict and evaluate credit risks in real time;

[0010] The internal structure and parameters of its model-free adaptive risk monitoring framework are optimized through an online learning algorithm.

[0011] Preferably, the step of constructing a model-free adaptive risk monitoring framework based on preprocessed data includes:

[0012] Determine the objectives and requirements of risk monitoring, including the types of risks to be monitored, risk sources, risk thresholds, and the expected monitoring accuracy and real-time performance;

[0013] Adopt a systematic approach to identify potential risks and conduct quantitative and qualitative assessments of identified risks;

[0014] A model-free adaptive risk monitoring framework is established according to the causal relationship between risk factors, and a model-free adaptive control algorithm is integrated into the model-free adaptive risk monitoring framework.

[0015] Preferably, the calculation formula of the pseudo partial derivative is as follows:

[0016] PPD(k)=Δy(k) / Δu(kd)

[0017] Among them, PPD(k) represents the pseudo partial derivative at the kth moment, Δy(k) represents the change in the system output at the kth moment, Δu(kd) represents the change in the system input at the kdth moment, and d is the delay time.

[0018] Preferably, the step of dynamically adjusting the risk monitoring strategy based on the degree of impact includes:

[0019] Use model-free adaptive control algorithms to monitor risk signals in credit operations in real time;

[0020] Categorize and assess the impact of captured risk signals;

[0021] Conduct risk analysis and trigger early warning mechanisms based on the results of risk analysis;

[0022] According to the type of risk and the degree of impact, the corresponding risk response strategy is formulated, and the model-free adaptive control algorithm is optimized according to the newly captured risk signals and the implementation results of the response strategy.

[0023] Preferably, a model-free adaptive prediction method is used to fill in the missing values ​​of personal credit data.

[0024] Preferably, the step of optimizing the internal structure and parameters of the model-free adaptive risk monitoring framework through an online learning algorithm comprises:

[0025] According to the data analysis results, determine the internal parameters that need to be optimized, and adjust the values ​​of internal parameters based on historical data and business logic;

[0026] Risk threshold adjustment: Risk thresholds are adjusted based on historical overdue rates and expected overdue rate targets;

[0027] Model weight adjustment: In machine learning-based models, risk prediction is optimized by adjusting feature weights. Weight adjustment is based on the importance of the features, i.e., W = g(I), where g is a function determined by the importance of the features.

[0028] Use the adjusted parameters to conduct simulation tests to evaluate their impact on risk prediction and delinquency rate. If the resulting effect does not meet the preset conditions, continue to adjust the parameters.

[0029] Preferably, in the model-free adaptive control algorithm, the output of the controller is usually calculated based on the input of the system and the PPD, expressed as u(k)=u(k-1)+K*e(k)*PPD(k)

[0030] Where u(k) represents the output of the controller at the kth iteration, u(k-1) represents the output of the controller at the previous iteration, K represents the controller gain, e(k) represents the error at the current moment, and PPD(k) represents the PPD value at the current moment.

[0031] Preferably, in the model-free adaptive control algorithm, the parameters of the controller are updated by the gradient descent method, and the gradient descent method is used to update the PPD value and the control rate. The general form of the update formula is:

[0032]

[0033]

[0034] Among them, η is the learning rate, which indicates the step size of parameter update; and are the gradients of the error function with respect to PPD and control rate, respectively.

[0035] Preferably, in the model-free adaptive control algorithm, the value function is used to evaluate the quality of the current state and control action, and the general form is:

[0036] J = Σ[e(t)2]

[0037] Among them, Σ represents the summation operation, and e(t)2 represents the square of the control action, which is used to measure the cost of the control action.

[0038] The credit risk monitoring method based on model-free adaptive control provided by the present invention not only improves the robustness and adaptability of the control system, but also significantly reduces the dependence on prior knowledge of the system, making the control strategy more flexible and efficient. In the future, with the further development of technologies such as big data and artificial intelligence, the model-free adaptive control method is expected to realize intelligent and adaptive control needs in more fields, and promote continuous progress in fields such as industrial automation and financial risk management. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 A flowchart of a credit risk monitoring method based on model-free adaptive control is provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0040] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0041] like Figure 1 FIG. 1 is a flowchart of a credit risk monitoring method based on model-free adaptive control provided by an embodiment of the present invention, wherein the method comprises:

[0042] Acquire personal credit data, preprocess the personal credit data, and obtain preprocessed data.

[0043] In this step, the original personal credit data is fully cleaned and preprocessed. This step includes data deduplication, outlier processing, and missing value identification, aiming to build a high-quality, noise-free data set to provide a solid foundation for subsequent analysis and modeling.

[0044] A model-free adaptive risk monitoring framework is constructed based on preprocessed data, and the model-free adaptive risk monitoring framework includes a dynamic learning module and a pseudo partial derivative calculation module.

[0045] In this step, a model-free adaptive risk monitoring framework is constructed using preprocessed personal credit data. The core of the framework is a dynamic learning module, which learns and updates the dynamic characteristics of credit data in real time based on the model-free adaptive control principle. In addition, the framework also integrates a pseudo-partial derivative calculation module for subsequent adaptive adjustment and optimization.

[0046] The steps to build a model-free adaptive risk monitoring framework include:

[0047] Clarify goals and needs:

[0048] First, it is necessary to clarify the objectives and requirements of risk monitoring. This includes determining the risk types to be monitored, risk sources, risk thresholds, and the expected monitoring accuracy and real-time performance. This step is the basis for building a risk monitoring framework and provides guidance for subsequent steps.

[0049] Risk identification and assessment:

[0050] 1. Use systematic methods (such as ABCD risk identification methods and tools, brainstorming, flowchart method, Delphi method, historical data analysis method, etc.) to identify potential risks.

[0051] 2. Comprehensively sort out and identify the risks that the company may face during its operations, including market risk, credit risk, operational risk, etc.

[0052] 3. Conduct quantitative and qualitative assessments of identified risks to determine the size, probability of occurrence, and potential impact of the risks; formulate risk assessment standards based on factors such as the probability of occurrence, degree of impact, and duration of the risks, and conduct quantitative assessments of the risks.

[0053] Build a model-free adaptive control algorithm:

[0054] In the model-free adaptive risk monitoring framework, the core is the model-free adaptive control algorithm. This algorithm does not rely on the mathematical model of the controlled system, but implements the design of the control law based on input and output data. The specific design steps may include:

[0055] Dynamic linearization: Convert a nonlinear system into a linear system so that linear control theory can be applied.

[0056] Parameter estimation: Use input and output data to estimate the time-varying parameters of the system (such as the PPD parameters mentioned below).

[0057] A model-free adaptive control (MFAC) algorithm is designed based on the above model.

[0058] Building a risk monitoring model:

[0059] After identifying the risk type and assessing the risk size, a risk monitoring model needs to be established based on the causal relationship between risk factors. This model should be able to reflect the risk status in real time and issue a warning in a timely manner when the risk exceeds the acceptable level.

[0060] Integrate model-free adaptive control algorithm with risk monitoring model:

[0061] The designed model-free adaptive control algorithm is integrated into the risk monitoring model to achieve the integration of monitoring and control. In this way, when the risk status changes, the control algorithm can adaptively adjust the control strategy to maintain the risk within an acceptable range.

[0062] Testing and Optimization:

[0063] 1. Testing: Test the constructed model-free adaptive risk monitoring framework to verify its effectiveness and accuracy.

[0064] 2. Optimization: Optimize the framework based on the test results, including adjusting algorithm parameters, improving risk identification and assessment methods, etc.

[0065] Deployment and Implementation:

[0066] The optimized model-free adaptive risk monitoring framework is deployed in the actual production environment, and corresponding operation specifications and maintenance plans are formulated. At the same time, a fast and effective emergency response mechanism is established to ensure timely response and mitigate losses when risk events occur.

[0067] Continuous monitoring and improvement:

[0068] Continuously monitor the implemented risk monitoring measures and regularly report the risk management effects. Continuously improve and optimize risk management strategies and methods based on changes in the internal and external environment and lessons learned from risk management practices.

[0069] Pseudo Partial Derivative (PPD) Principle and Formula

[0070] Pseudo partial derivative is an important concept in model-free adaptive control. It is used to describe the dynamic response relationship between system output and control input. In credit risk monitoring, we can regard credit data as the input of the system and risk warning signals as the output of the system. Pseudo partial derivative reflects the degree of influence of changes in credit data on risk warning signals.

[0071] The calculation formula of pseudo partial derivative is as follows:

[0072] PPD(k)=Δy(k) / Δu(kd)

[0073] Among them, PPD(k) represents the pseudo partial derivative at the kth moment, Δy(k) represents the change in the system output at the kth moment, Δu(kd) represents the change in the system input at the kdth moment, and d is the delay time.

[0074] Obtain the credit data to be processed, extract the risk characteristics of the credit data in real time through the dynamic learning module, use the pseudo partial derivative calculation module to calculate the impact of the credit data change on the risk warning signal, and dynamically adjust the risk monitoring strategy based on the impact, and the impact is represented by the pseudo partial derivative.

[0075] In this step, in the model-free adaptive risk monitoring framework, the risk characteristics of credit data are extracted in real time through the dynamic learning module. At the same time, the pseudo partial derivative calculation module is used to calculate the impact of credit data changes on risk warning signals (i.e., pseudo partial derivatives). According to the real-time changes of pseudo partial derivatives, the adaptive adjustment module dynamically adjusts the risk monitoring strategy to adapt to the complexity and uncertainty of credit data.

[0076] The specific steps for dynamic adjustment of credit risk monitoring are as follows:

[0077] Risk signal capture:

[0078] Real-time monitoring: Using model-free adaptive algorithms to monitor risk signals in credit business in real time, such as overdue payments and defaults;

[0079] Anomaly recognition: Identify abnormal risk signals through automatic detection by setting thresholds or algorithms.

[0080] Risk analysis assessment:

[0081] Risk classification: classify the captured risk signals, such as by risk type, risk level, etc.;

[0082] Impact assessment: Assess the specific impact of risk signals on credit business, including potential losses, changes in customer credit status, etc.

[0083] Strategy Adjustment:

[0084] Early warning mechanism: Based on the risk analysis results, the early warning mechanism is triggered to issue risk warnings to relevant personnel.

[0085] Strategy formulation: Develop corresponding risk response strategies based on the risk type and impact, such as adjusting credit limits and strengthening collection efforts.

[0086] Algorithm optimization: Based on the newly captured risk signals and the effectiveness of response strategies, the model-free adaptive algorithm is optimized to improve the accuracy and real-time performance of risk prediction.

[0087] Implementation and Monitoring:

[0088] Strategy implementation: Put the formulated risk response strategy into practice, such as adjusting customer credit ratings, strengthening risk control measures, etc.

[0089] Effect monitoring: Continuously monitor the implementation effect of risk response strategies, including the reduction of risk signals and business losses.

[0090] Feedback and iteration:

[0091] Effectiveness feedback: Feedback monitoring results to the risk management department and relevant personnel to understand the effectiveness of the strategy.

[0092] Iterative optimization: Based on feedback, the risk monitoring framework and response strategies are iteratively optimized to adapt to changes in credit risks and business development needs.

[0093] Build risk prediction models to predict and evaluate credit risks in real time.

[0094] In this step, advanced model-free adaptive prediction methods (such as time series-based prediction algorithms) are used to complete the missing values ​​that may exist in the credit data. Based on data completion, combined with pseudo partial derivative information, a risk prediction model is constructed to predict and evaluate credit risk in real time. The introduction of pseudo partial derivatives helps the model to more accurately capture the changing trend of credit data, thereby improving the accuracy of risk prediction.

[0095] The internal structure and parameters of its model-free adaptive risk monitoring framework are optimized through an online learning algorithm.

[0096] In this step, as credit data continues to accumulate and update, the model-free adaptive risk monitoring framework continuously optimizes its internal structure and parameters through online learning algorithms. The optimization process includes improving the accuracy of risk feature extraction, enhancing the robustness of risk prediction models, and fine-tuning adaptive adjustment strategies. At the same time, the model is further optimized and adjusted according to the changes in pseudo-partial derivatives. Through iterative updates, the framework can gradually adapt to the complexity and uncertainty of credit data and improve the accuracy and timeliness of risk monitoring.

[0097] The model-free adaptive risk monitoring framework continuously optimizes its internal structure and parameters through online learning algorithms, including:

[0098] Determine the optimization parameters:

[0099] Based on the data analysis results, determine the internal parameters that need to be optimized, such as risk thresholds, model weights, etc.

[0100] Parameter adjustment:

[0101] Adjust the values ​​of internal parameters based on historical data and business logic.

[0102] For example, if delinquency rates rise, one could consider lowering the risk threshold so that more credit applications are considered high risk.

[0103] Formula and principle:

[0104] (1) Risk threshold adjustment: The risk threshold (T) can be adjusted based on the historical overdue rate (R) and the expected overdue rate target (T_target). A simple formula can be: T = f(R, T_target), where f is a function determined according to business logic.

[0105] (2) Model weight adjustment: In machine learning-based models, risk prediction can be optimized by adjusting feature weights (W). Weight adjustment can be performed based on the importance (I) of the feature, that is, W = g(I), where g is a function determined based on the importance of the feature.

[0106] Effect evaluation:

[0107] (1) Conduct simulation tests using the adjusted parameters to evaluate their impact on risk prediction and delinquency rate.

[0108] (2) If the effect is not satisfactory, return to step 2 and continue adjusting the parameters.

[0109] Continuous monitoring and optimization

[0110] (1) Real-time monitoring:

[0111] Deploy an optimized credit risk monitoring system to monitor changes in credit risk in real time.

[0112] (2) Feedback and Adjustment:

[0113] Based on real-time monitoring results, the system performance is regularly evaluated.

[0114] If new risk trends or patterns are discovered, internal parameters are adjusted promptly to optimize system performance.

[0115] Continuously collect and analyze new data to discover new influencing factors and optimization points. At the same time, introduce new technologies and methods, such as deep learning and neural networks, to further improve the accuracy and real-time performance of risk prediction.

[0116] The core of the invention is to use the model-free adaptive control principle and the concept of pseudo partial derivative to build a monitoring framework that can learn in real time, adaptively adjust and accurately predict credit risk. This framework does not rely on traditional credit risk models, but is based on the dynamic characteristics and inherent laws of credit data, and achieves real-time monitoring and early warning of credit risk through real-time learning and adaptive adjustment.

[0117] The introduction of pseudo partial derivatives enables the framework to more accurately capture the changing trends of credit data and the degree of impact on risk warning signals, thereby improving the accuracy and timeliness of risk monitoring. At the same time, combined with advanced data completion and prediction methods, the integrity and accuracy of the data set are ensured, providing a reliable basis for risk prediction.

[0118] Through continuous optimization and iterative updates, the early warning framework can gradually adapt to the complexity and uncertainty of credit data and improve the robustness and adaptability of risk monitoring.

[0119] In the MFAC algorithm, commonly used formulas and algorithms include:

[0120] Pseudo partial derivative (PPD) update formula:

[0121] PPD is usually updated through an iterative algorithm to reflect changes in the dynamic characteristics of the system. The specific update formula may vary depending on the specific credit and financial risk scenario, but the general form is:

[0122] PPD(k+1)=PPD(k)+ΔPPD

[0123] Wherein, PPD(k) represents the PPD value at the kth iteration, and ΔPPD represents the increment of PPD.

[0124] Controller output calculation:

[0125] In the MFAC algorithm, the controller output is usually calculated based on the system input and PPD. The specific calculation formula may vary depending on the algorithm, but the general form is:

[0126] u(k)=u(k-1)+K*e(k)*PPD(k)

[0127] Among them, u(k) represents the output of the controller at the kth iteration, u(k-1) represents the output of the controller at the previous iteration, K represents the controller gain, e(k) represents the error at the current moment (that is, the difference between the set value and the actual value), and PPD(k) represents the PPD value at the current moment.

[0128] Error minimization algorithm:

[0129] The goal of the MFAC algorithm is to minimize the error between the set value and the actual value. To achieve this goal, the algorithm usually uses optimization algorithms such as gradient descent to update the controller parameters (such as PPD and controller gain K) to gradually approach the optimal control strategy.

[0130] In the credit risk monitoring method based on model-free adaptive control (MFAC), the error minimization algorithm is a key link. The following is a supplement to the details of the formula terminology in the algorithm:

[0131] 1. Error Definition

[0132] Error (e(t)) refers to the deviation between the set value (r(t)) and the actual output value (y(t)). Its mathematical expression is:

[0133] e(t)=r(t)-y(t)

[0134] In credit risk monitoring, the set value can be set to a certain risk threshold or expected risk level, while the actual output value is the credit risk value calculated by the MFAC model.

[0135] 2. Error Minimization Algorithm

[0136] The goal of the error minimization algorithm is to make the actual output value of the system as close to the set value as possible by adjusting the parameters of the controller (such as pseudo partial derivative PPD and control rate, etc.), that is, to minimize the error e(t).

[0137] Gradient Descent:

[0138] Gradient descent is a commonly used error minimization algorithm. It gradually approaches the optimal solution by calculating the gradient of the error function with respect to the controller parameters and updating the parameters in the opposite direction of the gradient.

[0139] In the MFAC algorithm, the gradient descent method can be used to update parameters such as PPD value and control rate. The specific update formula may vary depending on the algorithm, but the general form is:

[0140]

[0141]

[0142] Among them, η is the learning rate, which indicates the step size of parameter update; and are the gradients of the error function with respect to PPD and control rate, respectively.

[0143] Value function:

[0144] In the MFAC algorithm, the value function is used to evaluate the current state and the quality of the control action. By maximizing the value function, the optimal control action can be selected.

[0145] The specific form of the value function may vary depending on the application scenario, but the general form is:

[0146] J = Σ[e(t)2]

[0147] Among them, Σ represents the summation operation, and e(t)2 represents the square of the control action, which is used to measure the cost of the control action.

[0148] Policy Update:

[0149] Policy update refers to updating the controller parameters through an optimization algorithm to gradually optimize the control strategy. In the MFAC algorithm, policy update is usually based on the gradient descent method or other optimization algorithms.

[0150] Through continuous iterative learning, the MFAC algorithm can gradually converge to the optimal control strategy, making the actual output value of the system as close to the set value as possible, thereby minimizing the error.

[0151] 3. Algorithm implementation details:

[0152] When implementing the error minimization algorithm, you need to pay attention to the following details:

[0153] Sample collection: A reasonable sample collection strategy needs to be designed to obtain sufficient sample data for algorithm training and testing.

[0154] Parameter initialization: Reasonable initial values ​​need to be set for parameters such as PPD value and control rate to ensure the convergence and stability of the algorithm.

[0155] Learning rate selection: The choice of learning rate has an important impact on the convergence speed and stability of the algorithm. Too large a learning rate may cause the algorithm to diverge, while too small a learning rate may cause the algorithm to converge too slowly.

[0156] Weight factor adjustment: The weight factor λ is used to balance the control accuracy and the cost of the control action. In practical applications, the value of the weight factor needs to be adjusted according to the specific scenario.

[0157] In summary, the error minimization algorithm in the credit risk monitoring method based on model-free adaptive control (MFAC) is a key link in realizing adaptive control. By introducing concepts such as gradient descent, value function and strategy update, real-time monitoring and early warning of credit risk can be achieved, providing financial institutions with effective risk management tools.

[0158] The following are two cases of the application of the credit risk monitoring method based on model-free adaptive control (MFAC) in actual business, which include specific data changes and are explained by taking XXXX Finance and XXXX Credit Company as examples:

[0159] Case 1: Personal credit risk assessment of a certain financial company

[0160] background:

[0161] XXXX Financial Company is a financial institution focusing on personal credit business. Previously, its non-performing loan ratio has been maintained at around 2.5%. In order to improve the accuracy and timeliness of credit risk assessment, the company decided to introduce the MFAC method for credit risk monitoring.

[0162] Application Process:

[0163] 1. Data collection and preprocessing: A certain financial company collected multi-dimensional data such as customers' personal credit records, income status, consumption behavior, etc., and carried out comprehensive cleaning and preprocessing.

[0164] 1. Build an MFAC risk monitoring model: Using the preprocessed data, the company built an MFAC-based credit risk monitoring model that can learn and update the dynamic characteristics of credit data in real time.

[0165] 2. Risk feature extraction and assessment: Through the MFAC model, the company extracted the risk features of credit data in real time and conducted accurate credit risk assessment.

[0166] 3. Risk warning and decision support: The MFAC model issues risk warnings in a timely manner when the credit risk exceeds the preset threshold, providing decision support for the company.

[0167] Effect:

[0168] After introducing the MFAC method, the accuracy and timeliness of credit risk assessment of a certain financial company have been significantly improved. Through real-time monitoring and early warning, the company successfully identified and avoided multiple potential credit risk events. Within 6 months after the introduction of the MFAC method, the non-performing loan ratio dropped from 2.5% to 1.8%, a decrease of 28%. This significant improvement not only reduced credit risk, but also improved business operation efficiency.

[0169] Case 2: Supply Chain Finance Risk Management of a Certain Credit Company

[0170] background:

[0171] A certain credit company is a financial institution focusing on supply chain finance business. Previously, the non-performing loan rate of its supply chain finance business has remained at around 3%. In order to improve the level of supply chain finance risk management, the company decided to introduce the MFAC method for credit risk monitoring.

[0172] Application Process:

[0173] 1. Data integration and analysis: A certain credit company integrated multi-dimensional data such as transaction data, inventory data, logistics data, etc. of various companies in the supply chain and conducted a comprehensive analysis.

[0174] 2. Build an MFAC risk monitoring model: Using the integrated data, the company built a supply chain financial risk monitoring model based on MFAC, which can monitor the credit risk status of each enterprise in the supply chain in real time.

[0175] 3. Risk warning and response: The MFAC model issues a risk warning in a timely manner when the credit risk of a company in the supply chain exceeds the preset threshold. The company takes corresponding risk response measures based on the warning information.

[0176] 4. Optimize risk management strategies: Through the real-time monitoring and early warning functions of the MFAC model, a certain credit company continuously optimizes its risk management strategies.

[0177] Effect:

[0178] After the introduction of the MFAC method, the supply chain financial risk management level of a certain credit company has been significantly improved. Within 12 months after the introduction of the MFAC method, the non-performing loan ratio dropped from 3% to 2%, a decrease of 33.3%. This significant improvement not only reduced credit risk, but also improved business operation efficiency and market competitiveness. At the same time, by optimizing risk management strategies, the company also provided more favorable credit conditions for high-quality enterprises and strengthened supervision and risk control of high-risk enterprises.

[0179] The above two cases demonstrate the practical application effect and value of the MFAC method in credit risk monitoring. By introducing the MFAC method, financial institutions can more accurately assess credit risks, issue risk warnings in a timely manner, and take corresponding risk response measures, thereby reducing the non-performing loan rate and improving business operation efficiency and market competitiveness.

[0180] The credit risk monitoring method based on model-free adaptive control provided by the present invention has the following advantages:

[0181] 1. High precision of data completion: Through deep learning methods, missing data can be automatically learned and predicted, which has higher precision than traditional data completion methods.

[0182] 2. Powerful nonlinear data processing capabilities: Compared with traditional single statistical methods, deep learning can automatically extract the nonlinear characteristics of data and is suitable for processing complex personal credit data.

[0183] 3. Strong large-scale processing capabilities: This method can be applied to large-scale personal credit data, and the prediction accuracy can be further improved as the data increases.

[0184] 4. High degree of automation: The entire process can be automated, eliminating most of the data missing and errors caused by manual intervention as much as possible.

[0185] 5. Large iterative optimization space. In the process of promoting and continuously optimizing the evaluation method, the evaluation method does not need to redesign the entire model structure (add or delete existing logic and architecture). It only needs to use various algorithms covered by relevant deep learning to retrain parameters, saving the cost of optimization iteration.

[0186] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A credit risk monitoring method based on model-free adaptive control, characterized in that: The method comprises: Acquire personal credit data, pre-process the personal credit data, and obtain pre-processed data; Building a model-free adaptive risk monitoring framework based on preprocessed data, wherein the model-free adaptive risk monitoring framework includes a dynamic learning module and a pseudo partial derivative calculation module; Obtaining credit data to be processed, extracting risk features of the credit data in real time through a dynamic learning module, calculating the degree of influence of changes in credit data on risk warning signals using a pseudo partial derivative calculation module, and dynamically adjusting risk monitoring strategies based on the degree of influence, wherein the degree of influence is represented by a pseudo partial derivative; Build risk prediction models to predict and evaluate credit risks in real time; The internal structure and parameters of its model-free adaptive risk monitoring framework are optimized through an online learning algorithm.

2. The credit risk monitoring method based on model-free adaptive control according to claim 1, characterized in that: The step of constructing a model-free adaptive risk monitoring framework based on preprocessed data includes: Determine the objectives and requirements of risk monitoring, including the types of risks to be monitored, risk sources, risk thresholds, and the expected monitoring accuracy and real-time performance; Adopt a systematic approach to identify potential risks and conduct quantitative and qualitative assessments of identified risks; A model-free adaptive risk monitoring framework is established according to the causal relationship between risk factors, and a model-free adaptive control algorithm is integrated into the model-free adaptive risk monitoring framework.

3. The credit risk monitoring method based on model-free adaptive control according to claim 1, characterized in that: The calculation formula of the pseudo partial derivative is as follows: PPD(k)=Δy(k) / Δu(kd) Among them, PPD(k) represents the pseudo partial derivative at the kth moment, Δy(k) represents the change in the system output at the kth moment, Δu(kd) represents the change in the system input at the kdth moment, and d is the delay time.

4. The credit risk monitoring method based on model-free adaptive control according to claim 2, characterized in that: The steps to dynamically adjust the risk monitoring strategy based on the impact level include: Use model-free adaptive control algorithms to monitor risk signals in credit operations in real time; Categorize and assess the impact of captured risk signals; Conduct risk analysis and trigger early warning mechanisms based on the results of risk analysis; According to the type of risk and the degree of impact, the corresponding risk response strategy is formulated, and the model-free adaptive control algorithm is optimized according to the newly captured risk signals and the implementation results of the response strategy.

5. The credit risk monitoring method based on model-free adaptive control according to claim 1, characterized in that: A model-free adaptive prediction method is used to fill the missing values ​​of personal credit data.

6. The credit risk monitoring method based on model-free adaptive control according to claim 1, characterized in that: The steps of optimizing the internal structure and parameters of its model-free adaptive risk monitoring framework through online learning algorithms include: According to the data analysis results, determine the internal parameters that need to be optimized, and adjust the values ​​of internal parameters based on historical data and business logic; Risk threshold adjustment: Risk thresholds are adjusted based on historical overdue rates and expected overdue rate targets; Model weight adjustment: In machine learning-based models, risk prediction is optimized by adjusting feature weights. Weight adjustment is based on the importance of the features, i.e., W = g(I), where g is a function determined by the importance of the features. Use the adjusted parameters to conduct simulation tests to evaluate their impact on risk prediction and delinquency rate. If the resulting effect does not meet the preset conditions, continue to adjust the parameters.

7. The credit risk monitoring method based on model-free adaptive control according to claim 2, characterized in that: In the model-free adaptive control algorithm, the output of the controller is usually calculated based on the system input and PPD, expressed as u(k)=u(k-1)+K*e(k)*PPD(k) Where u(k) represents the output of the controller at the kth iteration, u(k-1) represents the output of the controller at the previous iteration, K represents the controller gain, e(k) represents the error at the current moment, and PPD(k) represents the PPD value at the current moment.

8. The credit risk monitoring method based on model-free adaptive control according to claim 2, characterized in that: In the model-free adaptive control algorithm, the gradient descent method is used to update the parameters of the controller. The gradient descent method is used to update the PPD value and the control rate. The general form of the update formula is: Among them, η is the learning rate, which indicates the step size of parameter update; and are the gradients of the error function with respect to PPD and control rate, respectively.

9. The credit risk monitoring method based on model-free adaptive control according to claim 2, characterized in that: In the model-free adaptive control algorithm, the value function is used to evaluate the quality of the current state and control action, and its general form is: J = Σ[e(t)2] Among them, Σ represents the summation operation, and e(t)2 represents the square of the control action, which is used to measure the cost of the control action.

Images