Financial risk assessment method and device based on large language model

Abstract

The invention discloses a financial risk assessment method and device based on a large language model. The method comprises the following steps: acquiring historical data of a user and external scene variable data, and preprocessing the acquired historical data of the user; constructing an input vector according to the preprocessed historical data of the user and the external scene variable data; using a large language model to generate a user future behavior sequence based on a condition generation technology; adding a time dimension into the future behavior sequence of the user to obtain dynamic change data of the user behavior along with time t; and obtaining a risk feature mean value based on the dynamic change number of the user behavior along with the time t, and calculating a comprehensive risk score through weighted summation. According to the method, the large language model is applied to financial user behavior prediction, the future behavior sequence of the user under the external scene variable is generated through the condition generation technology, and the scientificity and accuracy of risk assessment are improved.

Description

Technical Field

[0001] The present invention relates to the field of financial risk control technology, and in particular to a financial risk assessment method and device based on a large language model. Background Art

[0002] Traditional risk assessment methods have the following shortcomings:

[0003] (1) Static issues

[0004] Traditional risk assessment models primarily rely on static data such as users' historical transaction records and credit scores, failing to reflect changes in user behavior in dynamic future scenarios. This static nature makes it difficult for models to capture the dynamic responses of user behavior to external variables such as economic fluctuations and interest rate adjustments. For example, in the event of an economic recession, a decline in a user's repayment ability cannot be accurately predicted. Furthermore, many existing models rely on fixed rules or linear regression, ignoring nonlinear behavioral trends in complex scenarios, resulting in significant bias in risk assessment results.

[0005] (2) Lack of behavior generation

[0006] Existing technologies lack the ability to generate sequences of future user behavior, making it difficult to predict future consumption or repayment behavior based on historical user data and scenario variables. This limitation makes it difficult for financial institutions to simulate changing trends in user behavior in specific scenarios, such as interest rate fluctuations or economic crises. Furthermore, existing models often use static assumptions, assuming that future user behavior will be consistent with past behavior. This assumption ignores the significant impact of the external environment on user decision-making, making it difficult to identify potential risk exposure points.

[0007] (3) Insufficient decision support

[0008] The results of traditional risk assessments are usually presented in the form of a single score or classification, lacking comparative analysis of risk changes under different scenarios. For example, existing technologies make it difficult to assess the differences in risk for users under economic recession and economic recovery scenarios. In addition, existing technologies are unable to optimize loan amounts, interest rates, or product pricing strategies based on risk assessment results. Financial institutions need to rely on experience-based decision-making, making it difficult to achieve data-driven precision strategy optimization. Even more seriously, due to the lack of comprehensive consideration of the external environment (such as economic indicators and market interest rate fluctuations), financial institutions' risk prevention and control and decision-making often lack systematic scenario support.

[0009] (4) Insufficient integration of multidimensional features

[0010] Existing technologies for analyzing user behavior typically rely on single-dimensional features (such as credit scores or transaction records), ignoring the importance of unstructured data such as user emotions and psychological characteristics. Furthermore, traditional models struggle to deeply integrate multidimensional user characteristics (such as consumption frequency, repayment ability, and psychological traits), resulting in a one-sided portrayal of user behavior. Ignoring the complex interactions between features further reduces the accuracy of risk assessment. For example, changes in consumption patterns and fluctuations in psychological characteristics may jointly influence a user's repayment behavior, but traditional models fail to model these interactions, making it difficult to provide a comprehensive user profile and accurate risk prediction. Summary of the Invention

[0011] The purpose of the present invention is to provide a financial risk assessment method and device based on a large language model to address the deficiencies in the existing technology.

[0012] To achieve the above objectives, in a first aspect, the present invention provides a financial risk assessment method based on a large language model, comprising:

[0013] Obtain the user's historical data and external scene variable data, and preprocess the acquired user's historical data;

[0014] Construct the input vector X based on the pre-processed user historical data and external scenario variable data:

[0015] X=[Xu,Xs]

[0016] Among them, Xu is the user's historical data, and Xs is the external scenario variable data;

[0017] Use a large language model to generate user future behavior sequences based on conditional generation technology:

[0018] Y=LLM(X|Condition)

[0019] Among them, Y is the generated user future behavior sequence, LLM(.) indicates the use of large language model for processing, and Condition is the condition for generating the user future behavior sequence;

[0020] Add the time dimension to the user's future behavior sequence to obtain the dynamic change data Yt of the user's behavior over time t:

[0021] Yt=LLM(X,t)

[0022] Based on the dynamic change data Yt of the user behavior over time t, the risk feature mean f is obtained i (y), and the comprehensive risk score R is calculated by weighted summation:

[0023]

[0024] Where n is the total number of risk features extracted from the user's future behavior sequence Y, w i is the weight of the i-th risk feature.

[0025] Furthermore, the risk feature mean f i (y) includes the mean default probability f1(y), the mean underpayment ratio f2(y) and the mean overconsumption risk f3(y).

[0026] Furthermore, it also includes:

[0027] The loan amount is adjusted based on the comprehensive risk score R, as follows:

[0028] Loan Adjustment=Base Loan×(1-R)

[0029] Among them, Loan Adjustment is the adjusted loan amount, and Base Loan is the basic loan amount.

[0030] Furthermore, it also includes:

[0031] The loan interest rate is adjusted based on the comprehensive risk score R, as follows:

[0032] Adjusted Rate=Base Rate+R×k

[0033] Among them, Adjusted Rate is the adjusted loan interest rate, Base Rate is the basic loan interest rate, and k is the loan interest rate adjustment factor.

[0034] Furthermore, the data preprocessing method is as follows:

[0035] Adjust the historical data of numerical users to the set range, as follows:

[0036]

[0037] Where x′ is the historical data of the numerical user after adjustment, x is the historical data of the numerical user before adjustment, min(x) is the lower limit of the set range, and max(x) is the upper limit of the set range;

[0038] Reduce the noise of the user's historical data through the moving average method;

[0039] Integrate structured data and unstructured data into a unified vector representation through embedding technology.

[0040] Furthermore, the user's historical data includes transaction records, repayment habits and credit scores.

[0041] Furthermore, the external scenario variable data includes economic growth rate, interest rate change data and market fluctuation data.

[0042] In a second aspect, the present invention provides a financial risk assessment device based on a large language model, comprising a storage medium and a processor, wherein the storage medium stores a computer program, and the computer program is used to implement the above method when executed by the processor.

[0043] Beneficial Effects: 1. This invention applies the Large Language Model (LLM) to financial user behavior prediction, generating a user's future behavior sequence under external scenario variables through conditional generation technology. It dynamically combines user historical data with external scenario variable data to generate logical prediction data with time series characteristics. Compared with traditional risk assessment methods that rely on static data, this invention makes the assessment dynamic and predictive.

[0044] 2. This patent proposes a scenario-based risk assessment method that conducts multi-scenario comparative analysis of user behavior in different economic environments (such as recession and stability) to quantify the impact of scenario variables on risk. By extracting risk characteristics such as default probability and underpayment ratio from user behavior, and combining it with a dynamic weighting mechanism to calculate a comprehensive risk score, it comprehensively reflects the user's risk level in multiple scenarios. It can also identify risk exposure points in time series and capture high-risk time periods, providing support for risk warnings for financial institutions.

[0045] 3. The present invention constructs a complete closed-loop design system from user behavior generation, scenario-based risk assessment to decision support, which improves the efficiency of the entire process of financial risk control, provides high-quality predictive behavior data for risk assessment, quantifies risk characteristics and outputs dynamic scores, and ultimately realizes dynamic adjustment of loan amounts, interest rates and other strategies through the decision support module, enhancing the intelligence and practicality of risk management. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is a flowchart of a financial risk assessment method based on a large language model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0047] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. These embodiments are implemented based on the technical solutions of the present invention. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0048] like Figure 1 As shown, an embodiment of the present invention provides a financial risk assessment method based on a large language model, comprising:

[0049] Obtain user historical data and external scenario variable data, and preprocess the acquired user historical data. User historical data includes transaction records, repayment habits, and credit scores. External scenario variable data includes economic growth rate, interest rate change data, and market fluctuation data. The above data preprocessing method is as follows:

[0050] Adjust the historical data of numerical users to the set range, as follows:

[0051]

[0052] Here, x′ represents the adjusted numeric user's historical data, x represents the unadjusted numeric user's historical data, min(x) represents the lower limit of the set range, and max(x) represents the upper limit of the set range. A moving average method is then used to reduce noise in the user's historical data. Finally, structured data (such as credit scores) and unstructured data (such as customer complaint records) are integrated into a unified vector representation using embedding technology.

[0053] Construct the input vector X based on the pre-processed user historical data and external scenario variable data:

[0054] X=[Xu,Xs]

[0055] Where Xu is the user's historical data, and Xs is the external scenario variable data. Specifically, the user's historical data Xu = [credit score, transaction frequency, average monthly repayment], and the external scenario variable data Xs = [economic growth rate, interest rate fluctuation, inflation rate].

[0056] Use a large language model to generate user future behavior sequences based on conditional generation technology:

[0057] Y=LLM(X|Condition)

[0058] Where Y is the generated user's future behavior sequence, LLM(.) indicates processing using a large language model, and Condition is the condition for generating the user's future behavior sequence, such as "reducing consumption by 10%" or "delaying repayment by one month."

[0059] Add the time dimension to the user's future behavior sequence to obtain the dynamic change data Yt of the user's behavior over time t:

[0060] Yt=LLM(X,t)

[0061] The dynamic change data Yt of the above user behavior over time t includes consumption amount, repayment ability and default probability, etc. Among them:

[0062] Consumption amount: Combine economic growth rate and inflation rate to predict user consumption trends;

[0063] Repayment ability: Simulate repayment changes based on user repayment history and scenario variables;

[0064] Default probability: Dynamically adjust the default probability based on the user's credit score and underpayment.

[0065] For example:

[0066] t = 1 (month): Consumption amount = 14,000 yuan, Repayment = 11,500 yuan, Default probability = 0.05;

[0067] t = 3 (months): Consumption = 13,000 yuan, Repayment = 10,000 yuan, Default probability = 0.12;

[0068] t = 6 (months): Consumption amount = 12,000 yuan, Repayment = 9,000 yuan, Probability of default = 0.15.

[0069] Based on the dynamic change data Yt of the user behavior over time t, the risk feature mean f of different time steps is obtained i (y), and the comprehensive risk score R is calculated by weighted summation:

[0070]

[0071] Where n is the total number of risk features extracted from the user's future behavior sequence Y, w i is the weight of the mean of the i-th risk feature. For example, w1 is 0.5, w2 is 0.3, and w3 is 0.2.

[0072] Specifically, the risk characteristic mean f i (y) includes the mean default probability f1(y), the mean underpayment ratio f2(y) and the mean overconsumption risk f3(y). Among them, the mean default probability f1(y), the mean underpayment ratio f2(y) and the mean overconsumption risk f3(y) need to be obtained through corresponding calculations. Among them, the default probability f at any time step is 1j (y) can be directly extracted from the dynamically changing data Yt, and the underpayment ratio f at any time step 2j (y) and overconsumption risk f 3j (y) is calculated as follows:

[0073]

[0074] Where j is the serial number of the time step. It should be noted that the above-mentioned amount due, actual repayment amount, consumption amount and income are all within the range of the corresponding time step.

[0075] Based on the example data given above, the following examples are given:

[0076] Mean default probability f1(y):

[0077]

[0078] Mean underpayment ratio f2(y):

[0079] t=1,

[0080] t=3,

[0081] t=6,

[0082]

[0083] Overconsumption risk 3j (y):

[0084] t=1,

[0085] t=3,

[0086] t=6,

[0087]

[0088] After calculating the comprehensive risk score R, the loan amount can be adjusted based on the comprehensive risk score R, as follows:

[0089] Loan Adjustment=Base Loan×(1-R)

[0090] Among them, Loan Adjustment is the adjusted loan amount, and Base Loan is the basic loan amount.

[0091] The loan interest rate can also be adjusted based on the comprehensive risk score R, as follows:

[0092] Adjusted Rate=Base Rate+R×k

[0093] Among them, Adjusted Rate is the adjusted loan interest rate, Base Rate is the basic loan interest rate, and k is the loan interest rate adjustment factor.

[0094] Furthermore, the present invention can predict the comprehensive risk score R under different external scenario variables by inputting data on them. This allows for multi-scenario comparative analysis of user behavior in different economic environments (e.g., recession versus stability) to quantify the impact of scenario variables on risk. Furthermore, by calculating the comprehensive risk score at each time step, risk exposure points can be identified, capturing high-risk time periods and providing support for risk early warning for financial institutions.

[0095] The above description is merely a preferred embodiment of the present invention. It should be noted that any other aspects not specifically described are considered prior art or common knowledge to those skilled in the art. Improvements and modifications may be made without departing from the principles of the present invention, and such improvements and modifications are also within the scope of protection of the present invention.

Claims

1. A financial risk assessment method based on a large language model, characterized in that: include: Obtain the user's historical data and external scene variable data, and preprocess the acquired user's historical data; Construct the input vector X based on the pre-processed user historical data and external scenario variable data: X=[Xu,Xs] Among them, Xu is the user's historical data, and Xs is the external scenario variable data; Use a large language model to generate user future behavior sequences based on conditional generation technology: Y=LLM(X|Condition) Among them, Y is the generated user future behavior sequence, LLM(.) indicates the use of large language model for processing, and Condition is the condition for generating the user future behavior sequence; Add the time dimension to the user's future behavior sequence to obtain the dynamic change data Yt of the user's behavior over time t: Yt=LLM(X,t) Based on the dynamic change data Yt of the user behavior over time t, the risk feature mean f is obtained i (y), and the comprehensive risk score R is calculated by weighted summation: Where n is the total number of risk features extracted from the user's future behavior sequence Y, w i is the weight of the i-th risk feature.

2. A financial risk assessment method based on a large language model according to claim 1, characterized in that: The risk characteristic mean f i (y) includes the mean default probability f1(y), the mean underpayment ratio f2(y) and the mean overconsumption risk f3(y).

3. The financial risk assessment method based on a large language model according to claim 1, characterized in that: Also includes: The loan amount is adjusted based on the comprehensive risk score R, as follows: Loan Adjustment=Base Loan×(1-R) Among them, Loan Adjustment is the adjusted loan amount, and Base Loan is the basic loan amount.

4. The financial risk assessment method based on a large language model according to claim 1, characterized in that: Also includes: The loan interest rate is adjusted based on the comprehensive risk score R, as follows: Adjusted Rate=Base Rate+R×k Among them, Adjusted Rate is the adjusted loan interest rate, Base Rate is the basic loan interest rate, and k is the loan interest rate adjustment factor.

5. The financial risk assessment method based on a large language model according to claim 1, characterized in that: The data preprocessing method is as follows: Adjust the historical data of numerical users to the set range, as follows: Where x′ is the historical data of the numerical user after adjustment, x is the historical data of the numerical user before adjustment, min(x) is the lower limit of the set range, and max(x) is the upper limit of the set range; Reduce the noise of the user's historical data through the moving average method; Integrate structured data and unstructured data into a unified vector representation through embedding technology.

6. The financial risk assessment method based on a large language model according to claim 1, characterized in that: The user's historical data includes transaction records, repayment habits and credit scores.

7. The financial risk assessment method based on a large language model according to claim 1, characterized in that: The external scenario variable data includes economic growth rate, interest rate change data and market fluctuation data.

8. A financial risk assessment device based on a large language model, comprising a storage medium and a processor, wherein the storage medium stores a computer program, characterized in that: When the computer program is executed by a processor, it is used to implement the method according to any one of claims 1 to 7.

Images