Method for batch construction of risk time sequence features and quantitative information risk adjustment

Through multi-channel data acquisition and advanced time series modeling technology, combined with dynamic Bayesian networks and fuzzy logic control, the problem that traditional risk management systems cannot respond and identify nonlinear changes in real time is solved, and the bank's accurate assessment of risks is achieved and rapid response, improving business risk management capabilities.

CN118917935AInactive Publication Date: 2024-11-08BOC CONSUMER FINANCE CO LTD

Patent Information

Application Number
CN202411163304.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-23
Publication Date
2024-11-08
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

Traditional risk management systems rely on static credit scores and financial status assessments, and cannot respond to changes in customer behavior and market environment in real time. The data quality and formats from different data sources vary greatly, making it difficult to effectively integrate and analyze, and cannot effectively identify customer behavior. nonlinear changes in .

Method used

Through multi-channel data acquisition and preprocessing, time series features are extracted using sliding windows, coefficients of variation, and time difference methods, combined with autoregression model, recurrent neural network and dynamic Bayesian network for modeling, dynamically adjust system parameters, and fuzzy logic control and reinforcement learning algorithms to optimize risk adjustment strategies to achieve real-time risk assessment and early warning.

Benefits of technology

Significantly improves the bank's perception of risk changes, can update customers' risk scores in real time, quickly respond to risk changes, reduce business risks, improve business approval accuracy, enhance the ability to respond to complex risk scenarios, and ensure stable and sustainable business development. .

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118917935A_ABST
    Figure CN118917935A_ABST
Patent Text Reader

Abstract

The invention provides a method for batch construction of risk time sequence features and quantitative information risk adjustment, relates to the technical field of information processing, and comprises the steps of data cleaning, denoising and standardization processing. Feature engineering and time sequence modeling: modeling time sequence features by using an autoregression model, an exponential smoothing model and a recurrent neural network model; risk quantification, risk adjustment and feedback optimization are adopted, multi-channel data acquisition is adopted, a bank can obtain dynamic data of business behaviors, financial conditions and external economic environments of clients in real time, a preprocessing algorithm and a self-adaptive data acquisition strategy ensure the accuracy and timeliness of the data, the perception ability of the bank to risk changes is remarkably improved, and the risk change is more accurate. In combination with the time sequence characteristics and the dynamic Bayesian network, the system can update the risk score of the customer in real time, and adjust the business strategy in real time through a fuzzy logic control algorithm, so that the bank can quickly respond to the risk change, and the business risk is reduced.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of information processing technology, and in particular to a method for batch constructing risk time series characteristics and quantifying information risk adjustment. Background Art

[0002] According to the interest rate pricing method, device, electronic device and storage medium disclosed in Chinese patent number "CN113870002A", the method includes determining the target user's base lending rate through a preset pricing formula based on the target user's lending information; determining the default rate curve corresponding to the target user based on a pre-built default risk prediction model; determining the risk adjustment factor corresponding to the target lending rate through the default rate curve and the preset default rate; determining the target lending rate of the target user according to the risk adjustment factor and the base lending rate. The method disclosed in the present invention can start from the actual business scenario, introduce option pricing theory, comprehensively consider the base interest rate of macro factors such as industry (market) and micro factors such as financial institution costs, and combine the borrower's interest rate risk adjustment factor on the basis of the base interest rate to quantitatively realize the real-time pricing of the lending rate.

[0003] The above patent documents and prior art have the following technical problems when used:

[0004] 1. Traditional risk management systems usually rely on static credit scores and financial status assessments. These methods cannot respond to changes in customer behavior and market environment in real time, resulting in banks’ delayed response when facing risks.

[0005] 2. Traditionally, data from different data sources have large differences in quality and format, high data noise, and are difficult to integrate and analyze effectively;

[0006] 3. Traditional models can often only capture linear features in the data and cannot effectively identify nonlinear changes in customer behavior. Summary of the invention

[0007] Technical issues solved

[0008] In view of the shortcomings of the prior art, the present invention provides a method for batch constructing risk time series characteristics and quantifying information risk adjustment, which solves the following problems:

[0009] 1. Traditional risk management systems usually rely on static credit scores and financial status assessments. These methods cannot respond to changes in customer behavior and market environment in real time, resulting in banks’ delayed response when facing risks.

[0010] 2. Traditionally, data from different data sources have large differences in quality and format, high data noise, and are difficult to integrate and analyze effectively;

[0011] 3. Traditional models can often only capture linear features in the data and cannot effectively identify nonlinear changes in customer behavior.

[0012] Technical Solution

[0013] To achieve the above objectives, the present invention is implemented through the following technical solutions: a method for adjusting risk time series characteristics and quantitative information risks in batches, the adjustment method comprising the following parts:

[0014] Data collection: Collect raw data from multiple channels including historical data, environmental variables, and behavioral pattern data, and then process them through data cleaning, denoising, and standardization;

[0015] Feature engineering: Use sliding window, coefficient of variation, and time difference methods to extract time series features, generate new statistical features through derived feature construction, and use principal component analysis and clustering algorithms for feature screening and dimensionality reduction;

[0016] Time series modeling: Use autoregressive models, exponential smoothing models, and recurrent neural network models to model time series characteristics and predict future risk characteristic change trends;

[0017] Risk quantification: Based on the output of the time series model, calculate the time series volatility, evaluate the system stability, and set the risk warning threshold;

[0018] Risk adjustment: Based on the quantitative results, the system parameters are adjusted in real time through a dynamic adjustment mechanism, and an early warning mechanism is triggered when specific characteristics exceed the threshold;

[0019] Feedback optimization: By real-time monitoring of model output and actual results, model parameters are adjusted regularly and the model is retrained.

[0020] Preferably, in the data collection step, the data collection strategy is dynamically adjusted according to the quality, relevance and real-time performance of the data source by using an adaptive data screening algorithm, and high-quality data sources with the greatest impact on time series characteristics are preferentially selected, while low-relevance or noise data sources are excluded; the data source set is:

[0021] S=*s1,s2,…,s n +

[0022] The corresponding feature correlation measure is R(s i ), the quality metric is Q(s i ), the real-time metric is T(s i ), then the data source selection weight W(s i )for:

[0023] W(s i )=α·R(s i )+β·Q(si )+γ·T(s i )

[0024] α, β, γ are weight coefficients, select data source S * satisfy:

[0025] S * =*s i ∣W(s i )≥θ+

[0026] θ is the set threshold.

[0027] Preferably, the feature engineering uses kernel methods and polynomial expansion to transform the original linear features into a higher-dimensional nonlinear feature space to capture time series relationships, and uses regularization techniques to prevent model overfitting;

[0028] The original linear features are mapped to a high-dimensional space using kernel methods and polynomial expansion to capture complex time series relationships. The original feature vector is:

[0029] X=,x1,x2,…,x m -

[0030] Then the eigenvector after polynomial expansion is:

[0031]

[0032] Through the kernel function K(x i ,x j ) calculates the nonlinear relationship between features, using any of the following kernel functions:

[0033] Linear kernel: K(x i ,x j )=x i ·x j

[0034] Polynomial kernel: K(x i ,x j )=(γx i ·x j +r) d

[0035] Gaussian kernel (RBF kernel): K(x i ,x j )=exp(-γ∥x i -x j ∥ 2 ).

[0036] Preferably, in the time series modeling step, a hybrid model method is used to combine the autoregressive model with the recurrent neural network model to improve the accuracy and stability of the prediction;

[0037] The hybrid model method improves the prediction accuracy and stability of time series by combining the autoregressive model (AR) with the recurrent neural network model (RNN):

[0038]

[0039] Y t is the value of the time series at time t, φ i is the autoregressive coefficient, ∈ t is the error term, and the recurrent neural network model predicts the output at the current moment by memorizing the state information at the previous moment:

[0040] h t =σ(W h h t-1 +W x X t +b)

[0041] Y t =W y h t +c

[0042] h t is the implicit state, W h ,W x ,W y is the weight matrix, b and c are bias terms, and the hybrid model combines the advantages of both. The prediction result can be obtained by weighted average:

[0043]

[0044] λ1 and λ2 are weighted coefficients, and RNN is a neural function.

[0045] Preferably, the risk quantification adopts a dynamic Bayesian network model to dynamically adjust the risk probability distribution according to the changes in time series characteristics, and calculate the risk exposure under different scenarios through multi-level Monte Carlo simulation to achieve accurate risk quantification and assessment;

[0046] The dynamic Bayesian network (DBN) model is used to dynamically adjust the risk probability distribution according to the changes in time series characteristics. The dynamic Bayesian network describes the probability dependency of time series data over time. Each node: X t represents the state of the system at time t, and the conditional probability distribution is:

[0047]

[0048] Representation Node The parent node set of . Multi-level Monte Carlo simulation is used to evaluate risk exposure under different scenarios:

[0049]

[0050] Where N is the number of simulations, is the risk exposure value in the i-th simulation.

[0051] Preferably, the risk quantification adopts a spatiotemporal joint analysis algorithm to combine time series features with geographic spatial data, and to quantify and evaluate the distribution and evolution trend of risks in different geographical locations and time periods by constructing a spatiotemporal risk model;

[0052] The spatiotemporal joint analysis algorithm combines time series features with geographic spatial data, and evaluates the risk distribution and evolution trends in different time and space dimensions by constructing a spatiotemporal risk model. Mathematical formula: Assuming that the time series data is X(t) and the geographic spatial data is S(x,y), the spatiotemporal joint analysis model can be expressed as:

[0053] R(t,x,y)=f(X(t),S(x,y))

[0054] By analyzing the distribution and changing trend of R(t,x,y), the risk exposure values ​​at different geographical locations and time points can be determined.

[0055] Preferably, the feedback optimization module compares the historical adjustment results with the actual risk performance through a reinforcement learning algorithm, and automatically adjusts the model parameters to enhance the system's adaptive ability and risk resistance;

[0056] Reinforcement learning algorithms are used to continuously optimize risk adjustment strategies, enabling the system to autonomously adapt to changing environments. Mathematical formula: The basic framework of reinforcement learning includes state s t 、Action a t , Reward t and state transitions t+1 , the goal of strategy π is to maximize the expected cumulative reward:

[0057]

[0058] γ is the discount factor, and the Q learning algorithm updates the Q value function Q(s t ,a t ) to approximate the optimal strategy:

[0059]

[0060] α is the learning rate.

[0061] Preferably, the risk adjustment uses a fuzzy logic control algorithm to handle the uncertainty and ambiguity of different features in the time series;

[0062] The fuzzy logic control algorithm handles the uncertainty and ambiguity in the time series characteristics, dynamically adjusts the system parameters and regulation strategies, and the fuzzy logic system consists of three parts: fuzzification, rule reasoning, and defuzzification. The fuzzification step converts the input variables into fuzzy sets. The membership function of the input variable X is:

[0063]

[0064] c and σ are the center and width of the fuzzy set A respectively. The rule reasoning part is based on fuzzy rules. If X is a high risk, then Y is an emergency response. The defuzzification step converts the fuzzy output into an accurate value using the weighted average method:

[0065]

[0066] Y: output after defuzzification, Input x is in the fuzzy set A i The membership degree in indicates that x belongs to the fuzzy set A i The degree of membership is a value between 0 and 1, y i : and fuzzy set A i The corresponding representative value, n: the number of fuzzy outputs, that is, the number of fuzzy sets involved in the defuzzification calculation.

[0067] Preferably, the feature engineering adopts a feature enhancement algorithm based on a generative adversarial network (GANs) to generate time series features through adversarial training of a generator and a discriminator;

[0068] The generative adversarial network consists of a generator G and a discriminator D. The goal is that the data generated by the generator can deceive the discriminator so that it cannot distinguish between real data and generated data. The loss function is:

[0069]

[0070] p data (x) is the distribution of real data, p z (z) is the distribution of noise, G(z) is the generated data, and the generative adversarial network continuously optimizes the adversarial process between the generator and the discriminator, and finally generates time series features that are highly similar to the real data.

[0071] Preferably, the early warning mechanism includes a multi-model early warning integration algorithm for weighted combination of early warning signals output by different models to generate a comprehensive early warning signal;

[0072] There are N models, and the warning signal of each model is S i , whose weighted combination is:

[0073]

[0074] λ i is the weight of model i, satisfying Early warning signals can dynamically adjust weights λ based on historical forecast accuracy i , realizing an adaptive early warning mechanism.

[0075] Beneficial Effects

[0076] The present invention provides a batch construction risk time series feature and quantitative information risk adjustment method.

[0077] It has the following beneficial effects:

[0078] 1. The present invention adopts multi-channel data collection, so that banks can obtain dynamic data on customers' business behavior, financial status and external economic environment in real time. The preprocessing algorithm and adaptive data collection strategy ensure the accuracy and timeliness of the data, and significantly improve the bank's ability to perceive risk changes. Combined with time series characteristics and dynamic Bayesian networks, the system can update customers' risk scores in real time and adjust business strategies instantly through fuzzy logic control algorithms to ensure that banks can quickly respond to risk changes and reduce business risks.

[0079] 2. By combining the autoregressive model (AR) and the recurrent neural network (RNN), the present invention enables banks to accurately capture the linear and nonlinear characteristics of customer behavior and predict future risk trends, effectively improving the accuracy of business approval and reducing the risk of decision-making errors.

[0080] 3. The system of the present invention continuously compares the model prediction results with the actual business performance, identifies the differences and makes optimization adjustments. By regularly retraining the model, the bank's risk management system can always adapt to market changes and maintain efficient operation. As the system is continuously optimized and its ability to respond to complex risk scenarios is enhanced, the bank can maintain stable and sustainable business development in the face of market fluctuations, providing strong guarantees for sound operations. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] Figure 1 It is the structure cloud diagram of the present invention;

[0082] Figure 2 It is the relationship diagram between the risk level and the economic behavior cycle of the present invention;

[0083] Figure 3 A relationship diagram between the business approval efficiency and the fluctuation of the amount of economic behavior of the present invention;

[0084] Figure 4 It is the customer economic behavior trend diagram of the present invention;

[0085] Figure 5 This is a data diagram of the present invention. DETAILED DESCRIPTION

[0086] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. Specific embodiment one:

[0088] like Figure 1 As shown, a batch construction risk time series characteristics and quantitative information risk adjustment method, the adjustment method includes the following parts:

[0089] Data collection: Collect raw data from multiple channels including historical data, environmental variables, and behavioral pattern data, and then process them through data cleaning, denoising, and standardization;

[0090] Feature engineering: Use sliding window, coefficient of variation, and time difference methods to extract time series features, generate new statistical features through derived feature construction, and use principal component analysis and clustering algorithms for feature screening and dimensionality reduction;

[0091] Time series modeling: Use autoregressive models, exponential smoothing models, and recurrent neural network models to model time series characteristics and predict future risk characteristic change trends;

[0092] Risk quantification: Based on the output of the time series model, calculate the time series volatility, evaluate the system stability, and set the risk warning threshold;

[0093] Risk adjustment: Based on the quantitative results, the system parameters are adjusted in real time through a dynamic adjustment mechanism, and an early warning mechanism is triggered when specific characteristics exceed the threshold;

[0094] Feedback optimization: By real-time monitoring of model output and actual results, model parameters are adjusted regularly and the model is retrained.

[0095] In the data collection step, the data collection strategy is dynamically adjusted according to the quality, relevance and real-time performance of the data source by using an adaptive data screening algorithm, giving priority to high-quality data sources that have the greatest impact on time series characteristics, and excluding low-relevance or noise data sources; the data source set is:

[0096] S=*s1,s2,…,s n +

[0097] The corresponding feature correlation measure is R(s i ), the quality metric is Q(s i ), the real-time metric is T(s i), then the data source selection weight W(s i )for:

[0098] W(s i )=α·R(s i )+β·Q(s i )+γ·T(s i )

[0099] α, β, γ are weight coefficients, select data source S * satisfy:

[0100] S * =*s i ∣W(s i )≥θ+

[0101] θ is the set threshold.

[0102] Feature engineering uses kernel methods and polynomial expansion to transform the original linear features into a higher-dimensional nonlinear feature space to capture time series relationships, and uses regularization techniques to prevent model overfitting;

[0103] The original linear features are mapped to a high-dimensional space using kernel methods and polynomial expansion to capture complex time series relationships. The original feature vector is:

[0104] X=,x1,x2,…,x m -

[0105] Then the eigenvector after polynomial expansion is:

[0106]

[0107] Through the kernel function K(x i ,x j ) calculates the nonlinear relationship between features, using any of the following kernel functions:

[0108] Linear kernel: K(x i ,x j )=x i ·x j

[0109] Polynomial kernel: K(x i ,x j )=(γx i ·x j +r) d

[0110] Gaussian kernel (RBF kernel): K(x i ,x j )=exp(-γ∥x i -x j ∥2 ).

[0111] In the time series modeling step, a hybrid model method is used to combine the autoregressive model with the recurrent neural network model to improve the accuracy and stability of the prediction;

[0112] The hybrid model method improves the prediction accuracy and stability of time series by combining the autoregressive model (AR) with the recurrent neural network model (RNN):

[0113]

[0114] Y t is the value of the time series at time t, φ i is the autoregressive coefficient, ∈ t is the error term, and the recurrent neural network model predicts the output at the current moment by memorizing the state information at the previous moment:

[0115] h t =σ(W h h t-1 +W x X t +b)

[0116] Y t =W y h t +c

[0117] h t is the implicit state, W h ,W x ,W y is the weight matrix, b and c are bias terms, and the hybrid model combines the advantages of both. The prediction result can be obtained by weighted average:

[0118]

[0119] λ1 and λ2 are weighted coefficients, and RNN is a neural function.

[0120] Risk quantification uses a dynamic Bayesian network model to dynamically adjust the risk probability distribution according to the changes in time series characteristics, and calculates the risk exposure under different scenarios through multi-level Monte Carlo simulation to achieve accurate risk quantification and assessment;

[0121] The dynamic Bayesian network (DBN) model is used to dynamically adjust the risk probability distribution according to the changes in time series characteristics. The dynamic Bayesian network describes the probability dependency of time series data over time. Each node: X t represents the state of the system at time t, and the conditional probability distribution is:

[0122]

[0123] Representation Node The parent node set of . Multi-level Monte Carlo simulation is used to evaluate the risk exposure formula under different scenarios:

[0124]

[0125] Where N is the number of simulations, is the risk exposure value in the i-th simulation.

[0126] Risk quantification uses a spatiotemporal joint analysis algorithm to combine time series features with geospatial data. By building a spatiotemporal risk model, it quantifies and evaluates the distribution and evolution trend of risks in different geographical locations and time periods.

[0127] The spatiotemporal joint analysis algorithm combines time series features with geographic spatial data, and evaluates the risk distribution and evolution trends in different time and space dimensions by constructing a spatiotemporal risk model. Mathematical formula: Assuming that the time series data is X(t) and the geographic spatial data is S(x,y), the spatiotemporal joint analysis model can be expressed as:

[0128] R(t,x,y)=f(X(t),S(x,y))

[0129] By analyzing the distribution and changing trend of R(t,x,y), the risk exposure values ​​at different geographical locations and time points can be determined.

[0130] The feedback optimization module uses a reinforcement learning algorithm to compare historical adjustment results with actual risk performance and automatically adjust model parameters to enhance the system's adaptive ability and risk resistance.

[0131] Reinforcement learning algorithms are used to continuously optimize risk adjustment strategies, enabling the system to autonomously adapt to changing environments. Mathematical formula: The basic framework of reinforcement learning includes state s t 、Action a t , Reward t and state transitions t+1 , the goal of strategy π is to maximize the expected cumulative reward:

[0132]

[0133] γ is the discount factor, and the Q learning algorithm updates the Q value function Q(s t ,a t ) to approximate the optimal strategy:

[0134]

[0135] α is the learning rate.

[0136] Risk adjustment uses fuzzy logic control algorithm to deal with the uncertainty and ambiguity of different characteristics in time series;

[0137] The fuzzy logic control algorithm handles the uncertainty and ambiguity in the time series characteristics, dynamically adjusts the system parameters and regulation strategies, and the fuzzy logic system consists of three parts: fuzzification, rule reasoning, and defuzzification. The fuzzification step converts the input variables into fuzzy sets. The membership function of the input variable X is:

[0138]

[0139] c and σ are the center and width of the fuzzy set A respectively. The rule reasoning part is based on fuzzy rules. If X is a high risk, then Y is an emergency response. The defuzzification step converts the fuzzy output into an accurate value using the weighted average method:

[0140]

[0141] Y: output after defuzzification, Input x is in the fuzzy set A i The membership degree in indicates that x belongs to the fuzzy set A i The degree of membership is a value between 0 and 1, y i : and fuzzy set A i The corresponding representative value, n: the number of fuzzy outputs, that is, the number of fuzzy sets involved in the defuzzification calculation.

[0142] Feature engineering uses a feature enhancement algorithm based on generative adversarial networks (GANs) to generate time series features through adversarial training of generators and discriminators;

[0143] The generative adversarial network consists of a generator G and a discriminator D. The goal is that the data generated by the generator can deceive the discriminator so that it cannot distinguish between real data and generated data. The loss function is:

[0144]

[0145] p data (x) is the distribution of real data, p z (z) is the distribution of noise, G(z) is the generated data, and the generative adversarial network continuously optimizes the adversarial process between the generator and the discriminator, and finally generates time series features that are highly similar to the real data.

[0146] The early warning mechanism includes a multi-model early warning integration algorithm, which is used to perform a weighted combination of early warning signals output by different models to generate a comprehensive early warning signal;

[0147] There are N models, and the warning signal of each model is S i, whose weighted combination is:

[0148]

[0149] λ i is the weight of model i, satisfying Early warning signals can dynamically adjust weights λ based on historical forecast accuracy i , realizing an adaptive early warning mechanism. Specific embodiment 2:

[0151] See attached Figure 2-5, we can see that the main business of the bank includes corporate business and personal business. Due to the different customer characteristics, economic behavior, external economic environment and other factors of the two types of business, the risk management is more difficult. Traditional risk management systems usually use static credit scoring and financial status assessment methods, which cannot respond to customer behavior and market changes in real time, resulting in a lag in the bank's response to risks, which in turn leads to an increase in the non-performing business rate. Therefore, the bank introduced this technical solution and applied it to the risk management of its corporate business and personal business. Data collection and cleaning The bank first collected multiple data sources including historical economic behavior records, customer financial status, enterprise operation data, market industry data, macroeconomic data, social media public opinion data, etc. through multi-channel data collection. These data not only cover the customer's historical behavior and financial information, but also include the market environment and social and economic dynamics, providing a solid data foundation for the comprehensive assessment of business risks. After data collection, the bank used preprocessing algorithms to clean and denoise the data, and the standardization process ensured the comparability and consistency between different data sources. In addition, the bank also dynamically adjusted the data source through an adaptive data collection strategy to ensure that the collected data always has high relevance and quality. Feature engineering and screening: In the feature engineering stage, the bank extracts a series of time series features that can reflect risk characteristics through sliding window analysis, cumulative change analysis, trend analysis and other methods of customer behavior data and market data. For example, by analyzing the historical data of customers' economic behavior, the bank can extract characteristics such as the periodicity of customers' economic behavior and the volatility of the amount of economic behavior. Then, the bank uses principal component analysis (PCA) and LASSO regression techniques to screen and reduce the extracted features, remove redundant information, and retain key risk indicators. Finally, the bank builds a feature set containing multiple highly expressive features for subsequent risk prediction and quantification. Time series modeling and prediction: The bank uses hybrid modeling technology, combined with autoregressive model (AR) and recurrent neural network (RNN), to model the extracted time series features. The AR model is used to capture linear features in business data, such as the periodic fluctuations of the amount of customers' economic behavior, while the RNN is used to capture nonlinear features, such as mutations and trend changes in customers' economic behavior. By combining the advantages of these two models, the bank can more accurately predict risk changes in the future and warn potential high-risk customers and businesses in advance. Risk quantification and adjustment: In the risk quantification process, the bank dynamically associates time series characteristics with risk probability distribution through dynamic Bayesian networks (DBN). The system can adjust the risk level of each customer or business project in real time according to the changes in time series characteristics. In addition, the bank simulates and analyzes risks under various economic scenarios through Monte Carlo simulation technology to evaluate the risk exposure of businesses under different scenarios. These quantitative results help banks make more accurate judgments in business decisions and also provide a basis for risk adjustment.During the adjustment process, the bank uses fuzzy logic control algorithms to adjust business strategies in real time according to the risk quantification results, such as interest rate adjustment, business quota control, etc., so as to effectively reduce risk exposure. Feedback optimization and continuous improvement: During the implementation process, the bank continuously monitors the performance of the system through the feedback optimization mechanism, compares the model prediction results with the actual business performance, identifies differences and makes optimization adjustments. The bank regularly retrains the model to adapt to new market changes and customer behavior patterns to ensure that the system remains efficient and stable in actual applications. Through this feedback loop, the bank's risk management system can be continuously optimized and its ability to respond to complex risk scenarios can be gradually enhanced. By implementing this technical solution, the bank's business risk management capabilities have been significantly improved. First, the bank can obtain and analyze a large amount of customer behavior data and market data in real time, so as to make more accurate judgments in the process of business approval and risk warning. Secondly, the hybrid modeling technology enables the bank to have higher accuracy in predicting business risks, effectively reducing the incidence of non-performing businesses. In addition, the dynamic risk quantification and adjustment mechanism enables the bank to respond quickly to market changes, adjust business strategies in a timely manner, and minimize business risks. In actual application, the bank's non-performing business rate has dropped by 30% compared with before implementation, and the business approval efficiency has increased by 25%, providing strong guarantees for the bank's stable operation.

[0152] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the statement "comprising a reference structure" do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0153] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for batch constructing risk time series characteristics and quantifying information risk adjustment, characterized in that: The adjustment method comprises the following parts: Data collection: Collect raw data from multiple channels including historical data, environmental variables, and behavioral pattern data, and then process them through data cleaning, denoising, and standardization; feature Engineering: Use sliding window, coefficient of variation, and time difference methods to extract time series features, generate new statistical features through derived feature construction, and use clustering algorithms for feature screening and dimensionality reduction; Time series modeling: Use autoregressive models, exponential smoothing models, and recurrent neural network models to model time series characteristics and predict future risk characteristic change trends; Risk quantification: Based on the output of the time series model, calculate the time series volatility, evaluate the system stability, and set the risk warning threshold; Risk adjustment: Based on the quantitative results, the system parameters are adjusted in real time through a dynamic adjustment mechanism, and an early warning mechanism is triggered when specific characteristics exceed the threshold; Feedback optimization: By real-time monitoring of model output and actual results, model parameters are adjusted regularly and the model is retrained.

2. A method for adjusting risk time series characteristics and quantitative information risks in batches according to claim 1, characterized in that: In the data collection step, the data collection strategy is dynamically adjusted according to the quality, relevance and real-time performance of the data source by using an adaptive data screening algorithm, and the high-quality data source with the greatest impact on the time series characteristics is preferentially selected, and low-relevance or noise data sources are excluded; the data source set is: S={s1,s2,…,s n } The corresponding feature correlation measure is R(s i ), the quality metric is Q(s i ), the real-time metric is T(s i ), then the data source selection weight W(s i )for: W(s i )=α·R(s i )+β·Q(s i )+γ·T(s i ) α, β, γ are weight coefficients, select data source S * satisfy: S * ={s i ∣W(s i )≥θ} θ is the set threshold.

3. A method for adjusting risk time series characteristics and quantitative information risks in batches according to claim 1, characterized in that: The feature engineering uses kernel methods and polynomial expansion to transform the original linear features into a higher-dimensional nonlinear feature space to capture time series relationships, and uses regularization techniques to prevent model overfitting; The original linear features are mapped to a high-dimensional space using kernel methods and polynomial expansion to capture complex time series relationships. The original feature vector is: X=[x1,x2,…,x m ] Then the eigenvector after polynomial expansion is: Through the kernel function K(x i ,x j ) calculates the nonlinear relationship between features, using any of the following kernel functions: Linear kernel: K(x i ,x j )=x i ·x j K(x i ,x j )=(γx i ·x j +r) d Gaussian kernel (RBF kernel): K(x i ,x j )=exp(-γ∥x i -x j ∥ 2 ).

4. A method for adjusting risk time series characteristics and quantitative information risks in batches according to claim 1, characterized in that: In the time series modeling step, a hybrid model method is used to combine the autoregressive model with the recurrent neural network model to improve the accuracy and stability of the prediction; The hybrid model method improves the prediction accuracy and stability of time series by combining the autoregressive model (AR) with the recurrent neural network model (RNN): Y t is the value of the time series at time t, φ i is the autoregressive coefficient, ∈ t is the error term, and the recurrent neural network model predicts the output at the current moment by memorizing the state information at the previous moment: h i =σ(W h h t-1 +W x X t +b) Y t =W y h t +c h t is the implicit state, W h ,W x ,W y is the weight matrix, b and c are bias terms, and the hybrid model combines the advantages of both. The prediction result can be obtained by weighted average: λ1 and λ2 are weighted coefficients, and RNN is a neural function.

5. A batch construction risk time series characteristics and quantitative information risk adjustment method according to claim 1, characterized in that: The risk quantification adopts a dynamic Bayesian network model to dynamically adjust the risk probability distribution according to the changes in time series characteristics, and calculates the risk exposure under different scenarios through multi-level Monte Carlo simulation to achieve accurate risk quantification and assessment; The dynamic Bayesian network (DBN) model is used to dynamically adjust the risk probability distribution according to the changes in time series characteristics. The dynamic Bayesian network describes the probability dependency of time series data over time. Each node: X t represents the state of the system at time t, and the conditional probability distribution is: Representation Node A collection of parent nodes, multi-level Monte Carlo simulation is used to evaluate risk exposure under different scenarios: Where N is the number of simulations, is the risk exposure value in the i-th simulation.

6. A batch construction risk time series characteristics and quantitative information risk adjustment method according to claim 1, characterized in that: The risk quantification adopts a spatiotemporal joint analysis algorithm to combine time series characteristics with geographic spatial data, and builds a spatiotemporal risk model to quantify and evaluate the distribution and evolution trend of risks in different geographical locations and time periods; The spatiotemporal joint analysis algorithm combines time series features with geographic spatial data, and evaluates the risk distribution and evolution trends in different time and space dimensions by constructing a spatiotemporal risk model: the time series data is X(t), the geographic spatial data is S(x,y), and the spatiotemporal joint analysis model can be expressed as: R(t,x,y)=f(X(t),S(x,y)) By analyzing the distribution and changing trend of R(t,x,y), the risk exposure values ​​at different geographical locations and time points can be determined.

7. A method for adjusting risk time series characteristics and quantitative information risks in batches according to claim 1, characterized in that: The feedback optimization module uses a reinforcement learning algorithm to compare historical adjustment results with actual risk performance and automatically adjust model parameters to enhance the system's adaptive ability and risk resistance. Reinforcement learning algorithms are used to continuously optimize risk adjustment strategies, enabling the system to autonomously adapt to the changing environment: The basic framework of reinforcement learning includes state s t 、Action a t , Reward t and state transitions t+1 , the goal of strategy π is to maximize the expected cumulative reward: γ is the discount factor, and the Q learning algorithm updates the Q value function Q(s t ,a t ) to approximate the optimal strategy: α is the learning rate.

8. A method for adjusting risk time series characteristics and quantitative information risks in batches according to claim 1, characterized in that: The risk adjustment adopts fuzzy logic control algorithm to deal with the uncertainty and ambiguity of different characteristics in time series; The fuzzy logic control algorithm handles the uncertainty and ambiguity in the time series characteristics, dynamically adjusts the system parameters and regulation strategies. The fuzzy logic system consists of three parts: fuzzification, rule reasoning, and defuzzification. The fuzzification step converts the input variables into fuzzy sets. The membership function of the input variable X is: c and σ are the center and width of the fuzzy set A respectively. The rule reasoning part is based on fuzzy rules. If X is a high risk, then Y is an emergency response. The defuzzification step converts the fuzzy output into an accurate value using the weighted average method: Y: output after defuzzification, Input x is in the fuzzy set A i The membership degree in indicates that x belongs to the fuzzy set A i The degree of membership is a value between 0 and 1, y i : and fuzzy set A i The corresponding representative value, n: the number of fuzzy outputs, that is, the number of fuzzy sets involved in the defuzzification calculation.

9. A method for adjusting risk time series characteristics and quantitative information risks in batches according to claim 1, characterized in that: The feature engineering adopts a feature enhancement algorithm based on a generative adversarial network (GANs) to generate time series features through adversarial training of a generator and a discriminator; The generative adversarial network consists of a generator G and a discriminator D. The goal is that the data generated by the generator can deceive the discriminator so that it cannot distinguish between real data and generated data. The loss function is: p data (x) is the distribution of real data, p z (z) is the distribution of noise, G(z) is the generated data, and the generative adversarial network continuously optimizes the adversarial process between the generator and the discriminator, and finally generates time series features that are highly similar to the real data.

10. A method for adjusting risk time series characteristics and quantitative information risks in batches according to claim 1, characterized in that: The early warning mechanism includes a multi-model early warning integration algorithm for weighted combination of early warning signals output by different models to generate a comprehensive early warning signal; There are N models, and the warning signal of each model is S i , whose weighted combination is: λ i is the weight of model i, satisfying Early warning signals can dynamically adjust weights λ based on historical forecast accuracy i , realizing an adaptive early warning mechanism.

Citation Information

Patent Citations

  • Long and short memory network and depth data cleaning-based financial time sequence prediction method, device and server

    CN110334881A

  • Credit risk scoring model training method and credit risk scoring method

    CN115829722A

  • Method for applying time sequence model to behavior variable

    CN116362860A

  • Digital asset market risk prediction system based on time sequence analysis

    CN116611696A

  • Dynamic optimization method and system for risk measurement model

    CN118134631A

Cited By

  • Big data driven compliance risk prediction and management method

    CN119151660A

  • Large container crane intelligent early warning system based on personnel behavior analysis

    CN119600783A

  • Agricultural product supply chain management method and system based on order data

    CN119693100A

  • Bank risk control data processing method and device, equipment and storage medium

    CN119941372A

  • Risk early warning method and system and electronic equipment

    CN120162213A