Second-hand car financial credit scoring method and system based on AI

Abstract

The invention discloses an AI-based second-hand car financial credit scoring method and system, and particularly relates to the field of artificial intelligence, and the method comprises the steps: S1, data calling, S2, preliminary screening, S3, model construction, S4, comprehensive judgment, S5, screening of credit clients, and S6, score updating. According to the method, three types of core data including credit records, vehicle estimation values and vehicle condition histories are called from a third-party data source through user authorization, credit scenes are identified to screen alternative customers, and a credit evaluation model library including a personal credit model, a vehicle risk model and a behavior fraud model is constructed for the alternative customers; each sub-model adopts a differentiation algorithm to adapt to different risk dimensions so as to construct a comprehensive feature vector, the comprehensive feature vector is used as the input of the nonlinear evaluation model, credit granting objects are screened, a credit comparison difference is obtained by comparing a secondary comprehensive credit score with an initial comprehensive credit score for the credit granting objects, and the credit granting objects are evaluated. And continuous monitoring of the second-hand car financial risk is realized.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to an AI-based method and system for credit scoring in the financial sector of used cars. Background Technology

[0002] The used car finance market is an important part of auto finance and consumer finance, and has shown a rapid growth trend in recent years. Compared with new car finance, used car finance faces more complex credit risk assessment challenges due to the dynamic decay of value, asymmetric vehicle condition information, and diversified credit qualifications of customer groups. Therefore, it is necessary to score the credit of used car finance to reduce the purchase risk of customers.

[0003] Traditional credit scoring methods mainly rely on strong financial data such as the applicant's central bank credit report, income certificate, and social security records. They also combine credit scoring models to statistically analyze historical data, select variables that are strongly correlated with default risk (such as age, income, and number of past delinquencies), assign weights to them, and generate a total score card.

[0004] However, it still has some shortcomings in actual use. First, the current credit assessment for used car finance is mainly based on the comparison and analysis of the applicant's traditional strong financial data such as the central bank's credit report and income certificate with the preset scoring model. However, credit risk is usually affected by the applicant's multi-dimensional social attributes, behavioral data and vehicle condition. The data dimensions of the existing assessment model are relatively simple and cannot provide comprehensive information support for the risk control decision of financial institutions. The accuracy of risk identification still needs to be improved.

[0005] Second, due to the limitations of the model's capabilities, risk control personnel need to rely on experience to investigate potential fraud or default risks one by one when they receive high-risk warnings. Because the model's interpretability is weak and it fails to incorporate dynamic risks such as vehicle value decay, the risk investigation and decision-making process may consume a lot of manpower, which is not conducive to the rapid approval and scale expansion of business. It is still necessary to further improve the efficiency and automation level of credit assessment. Summary of the Invention

[0006] In view of this, embodiments of the present invention provide an AI-based used car financial credit scoring method and system. With user authorization, a credit assessment model library is constructed, including a personal credit model, a vehicle risk model, and a behavioral fraud model. Each sub-model adopts differentiated algorithms to adapt to different risk dimensions, thereby screening candidate customers and creditworthy users, effectively solving the problems raised in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] S1: Data Retrieval: With user authorization, an access point is set up on the credit application page of the financial platform, allowing applicants to submit basic personal information and financial information, and authorizing the platform to retrieve credit records, vehicle valuations, and vehicle condition history from third-party data sources;

[0009] S2: Preliminary screening: Based on preset data processing rules, credit records, vehicle valuations, and vehicle condition history are retrieved from third-party data sources to identify credit scenarios, thereby initially screening applicants who have basic credit qualifications and marking them as candidate customers;

[0010] S3: Model Building: Retrieve the corresponding effective risk control features of the candidate customers, and then build a credit assessment model library. The credit assessment model library is used to store and run AI sub-models, including personal credit model, vehicle risk model and behavioral fraud model.

[0011] S4: Comprehensive Judgment: Integrate the output results of the AI ​​sub-models, and use a non-linear evaluation model to aggregate the probability of default, vehicle risk probability and behavioral fraud probability for comprehensive judgment, thereby predicting the probability of default risk.

[0012] S5: Screening Creditworthy Customers: Mapping the probability of default risk to a comprehensive credit score and automatically matching it with a preset risk level, thereby screening out creditworthy customers from the candidate customers;

[0013] S6: Score Update: Regularly retrieve credit records, vehicle valuations, and vehicle condition history from third-party data sources for creditworthy users and update their comprehensive credit scores. Trigger risk warnings for creditworthy users whose identities change abnormally.

[0014] This invention also provides an AI-based used car financial credit scoring system, comprising:

[0015] Data retrieval module: With user authorization, an access point is set up on the credit application page of the financial platform, allowing applicants to submit basic personal information and financial information, and authorizing the platform to retrieve credit records, vehicle valuations and vehicle condition history from third-party data sources;

[0016] Preliminary screening module: Based on preset data processing rules, it retrieves credit records, vehicle valuations, and vehicle condition history from third-party data sources to identify credit scenarios, and then preliminarily screens out applicants who have basic credit qualifications and marks them as candidate customers;

[0017] Model building module: retrieves the corresponding effective risk control features of candidate customers, and then builds a credit assessment model library. The credit assessment model library is used to store and run AI sub-models, including personal credit model, vehicle risk model and behavioral fraud model.

[0018] Comprehensive judgment module: It integrates the output results of the AI ​​sub-models and uses a non-linear evaluation model to aggregate its own default probability, vehicle risk probability, and behavioral fraud probability for comprehensive judgment, thereby predicting the default risk probability;

[0019] Secondary screening module: Maps the probability of default risk to a comprehensive credit score and automatically matches it with a preset risk level, thereby selecting creditworthy customers from the candidate customers;

[0020] Credit score update module: Regularly retrieves credit records, vehicle valuations, and vehicle condition history from third-party data sources for creditworthy users and updates their comprehensive credit score. Triggers risk warnings for creditworthy users whose identities change abnormally.

[0021] The technical effects and advantages of this invention are as follows:

[0022] 1. This invention retrieves risk signals based on the applicant's basic personal information and financial information. With user authorization, it retrieves three core data types from third-party data sources: credit records, vehicle valuation, and vehicle condition history. It also establishes a process of feature classification, risk signal matching, and scenario frequency statistics to identify credit scenarios and screen potential customers. This allows for accurate identification of the applicant's main credit scenarios and provides a more comprehensive basis for preliminary screening.

[0023] 2. The present invention further calculates interaction terms on the output probabilities of the AI ​​sub-model (probability of default, probability of vehicle risk, and probability of behavioral fraud) to construct a comprehensive feature vector, and uses this as the input of the nonlinear evaluation model. This method can explicitly capture the synergistic and enhancing effects between different risk dimensions, thereby achieving better risk prediction accuracy than simple linear weighting or direct feature splicing.

[0024] 3. This invention classifies risk levels into low, medium, and high risk based on comprehensive credit scoring, and clearly selects medium- and low-risk candidates as creditworthy customers. This effectively controls risk while maximizing the value of high-quality customers, achieving a balance between risk and business returns. Furthermore, for creditworthy users, credit records, vehicle valuations, and historical vehicle condition data are periodically retrieved from third-party data sources to recalculate their comprehensive credit scores. By comparing the secondary comprehensive credit score with the initial comprehensive credit score, the credit difference is obtained. This approach overcomes the limitations of traditional static credit scoring in addressing dynamic changes in user credit status and vehicle condition, enabling continuous monitoring of used car financial risks. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the overall structure of the present invention.

[0026] Figure 2 This is a flowchart of the system operation of the present invention.

[0027] Figure 3 This is a flowchart of the AI ​​sub-model output of the present invention. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] like Figure 1 - Figure 3 The illustrated AI-based used car finance credit scoring method includes the following steps:

[0030] S1: Data Retrieval: With user authorization, an access point is set up on the credit application page of the financial platform, allowing applicants to submit basic personal information and financial information, and authorizing the platform to retrieve credit records, vehicle valuations, and vehicle condition history from third-party data sources.

[0031] In the specific implementation of the above scheme, a functional module, such as a "Submit Personal Information" button or link, is integrated into the credit application page of the financial platform to guide applicants to upload their basic personal information and financial information. The basic personal information includes, but is not limited to, the applicant's identity information, occupation, and address. The financial information includes, but is not limited to, proof of income and debt status. After the applicant submits the information, the authorized platform (such as a credit reporting agency, vehicle data platform, or operator) retrieves credit records, vehicle valuations, and vehicle condition history based on the applicant's personal information. The credit records are obtained from the central bank's credit reporting system or licensed credit reporting agencies, including historical credit records and debt-to-income ratios. The vehicle valuation is obtained from a professional used car valuation platform, including the target vehicle's real-time market valuation price, residual value curve, and loan-to-value ratio. The vehicle condition history is obtained from a vehicle history data platform (such as a maintenance platform or insurance company database), including the vehicle's VIN code verification information, historical maintenance records, accident records, and actual mileage.

[0032] S2: Preliminary screening: Based on preset data processing rules, credit records, vehicle valuations, and vehicle condition history are retrieved from third-party data sources to identify credit scenarios, and applicants with basic credit qualifications are preliminarily screened and marked as candidate customers.

[0033] It should be explained that the preset data processing rules include missing value handling, outlier handling, and standardization. For example, for numerical features (such as mileage), missing values ​​can be filled using the median. Outlier handling can specifically use the interquartile range method for identification. For example, values ​​exceeding the range of [Q1-1.5IQR, Q3+1.5IQR] are considered outliers and are truncated or replaced with boundary values. Standardization is specifically achieved through Z-score standardization or Min-Max normalization to ensure that effective risk control features are on the same scale, which facilitates model processing.

[0034] In this embodiment, the specific identification steps for the credit scenario are as follows:

[0035] A1: Classify and analyze credit records, vehicle valuations, and vehicle condition history, and identify key risk signals from the characteristics of the classifications. Key risk scenarios include, but are not limited to, high-risk signals, low-risk signals, and medium-risk signals.

[0036] The purpose of the above feature classification is to associate multi-dimensional credit records, vehicle valuation, and vehicle condition history with specific risk types in order to more accurately identify the applicant's potential risk profile and support subsequent risk control decisions. For example, feature classification can be carried out using business rules or decision tree paths.

[0037] It needs to be explained that the analysis of effective risk control features in business rules is based on a series of typical credit scenarios and their corresponding characteristics predefined by risk control experts. When the system extracts credit records, vehicle valuations, and vehicle condition history, it directly matches them with these predefined scenario templates. For example, if the number of overdue payments is greater than 3 and the loan-to-value ratio is greater than or equal to 80% based on credit records, vehicle valuations, and vehicle condition history, then the key risk signal is identified as a high-risk signal. The decision tree path method recursively divides the data feature space to form a discrimination path from the root node to the leaf node. Each path represents a logical combination of a set of feature conditions and the corresponding decision result. For example, based on historical sample data, the optimal tree discrimination structure is constructed. The root node is initially divided according to the number of overdue payments greater than 3 in the past two years. Subsequent nodes are expanded according to conditions such as debt-to-income ratio > 55% and loan-to-value ratio > 80%, forming a complete discrimination path. The path "number of overdue payments, debt-to-income ratio, and loan-to-value ratio" are then combined to form a risk discrimination rule. The three features in the above path are identified as key dimensions that jointly affect risk judgment. Finally, the extracted rule is used to identify specific risk signals.

[0038] A2: Count the number of key risk signals identified and match each risk signal with the typical risk signals of different credit scenarios in the scenario rule base to obtain the credit scenario matched by each risk signal.

[0039] It should be added that credit scenarios include, but are not limited to, high-risk debt scenarios, high-risk vehicle scenarios, fraud scenarios, high-quality customer scenarios, and high-value scenarios.

[0040] Applying to the above embodiments, different credit risk categories in the scenario rule base usually have specific risk signals to help risk control personnel better understand and make decisions. For example, signals that are usually present in high-risk debt scenarios include "more than 3 overdue payments", and signals that are usually present in high-risk vehicle scenarios include "loan-to-value ratio > 80%".

[0041] A3: Analyze the frequency of occurrence of each credit scenario that matches each risk signal, and then take the scenario with the highest frequency of occurrence as the applicant's main credit scenario.

[0042] It should be added that multiple credit scenarios can be identified based on effective risk control characteristics. At this time, by statistically analyzing the frequency of each scenario, the most accurate scenario assessment can be made for the applicant, thereby making differentiated decisions such as approval, rejection, increasing interest rates or reducing credit limits.

[0043] Furthermore, if the applicant's primary credit scenario is positive, the applicant will be selected and marked as a candidate customer; if the applicant's primary credit scenario is negative, the applicant's used car sales will be automatically rejected.

[0044] It should be explained that positive credit scenarios are combinations of characteristics or business situations that can significantly reduce the applicant's overall default risk or indicate that they have a strong ability and willingness to repay, specifically high-quality customer scenarios and high-value scenarios. Negative credit scenarios are scenarios that increase the applicant's overall default risk or indicate that they have serious problems with their repayment ability and willingness to repay, specifically high-risk debt scenarios, high-risk vehicle scenarios, and fraud scenarios.

[0045] S3: Model Building: Retrieve the corresponding effective risk control features for candidate customers, and then build a credit assessment model library. The credit assessment model library is used to store and run AI sub-models, including personal credit model, vehicle risk model and behavioral fraud model.

[0046] In this embodiment, it should be specifically explained that effective risk control features include, but are not limited to, loan-to-value ratio, delinquency score, vehicle valuation volatility, vehicle depreciation rate, identity matching degree, and debt default rate. The loan-to-value ratio is obtained by extracting the loan amount and the vehicle's market valuation from credit records and vehicle valuations, and then comparing them. A higher loan-to-value ratio indicates a greater risk exposure for the financial institution; in the event of default, the disposal of the vehicle may not be able to cover the loan principal and interest. Vehicle valuation volatility is obtained by comparing the difference between the highest and lowest valuations over the past six months with the average valuation over the past six months. Higher volatility indicates poorer vehicle value stability. Price difference losses are easily incurred during disposal; vehicle depreciation rate is the difference between the value 1 and the vehicle's market appraisal price and guide price, reflecting the degree of physical and market damage to the vehicle; overdue score is obtained by multiplying the overdue number score and the overdue days score, where the overdue number score and overdue days score are obtained by matching according to the scoring table; identity matching degree is obtained by comparing the geographical distance of ID card address and IP address with the attenuation radius and subtracting it from the value 1. The attenuation radius can be set according to business, such as 300 kilometers (within the province) or 800 kilometers (across provinces), and then the difference result is compared with 0. The closer the actual geographical distance, the higher the identity matching degree.

[0047] It should be added that the scoring table can specifically be a scoring table for the number of overdue payments and a scoring table for the number of overdue days. For example, when the number of overdue payments is 0, the score can be 10; when the number of overdue payments is 1-2, the score can be 8; when the number of overdue payments is 3-5, the score can be 5; and when the number of overdue payments is greater than 5, the score is 1. The scoring table for the number of overdue days can be divided according to the number of days. When the number of overdue days is 0, the score is 10; when the number of overdue days is less than one month, the score is 7; when the number of overdue days is less than two months, the score is 4; and when the number of overdue days is greater than two months, the score is 1.

[0048] It should be explained that, because the model requires a large number of samples labeled "default" or "non-default" for supervised learning, it is necessary to extract the debt default rate. The specific extraction steps are as follows:

[0049] Establish a time window to observe whether the customer defaults, such as 12 months or 24 months after the loan is disbursed;

[0050] Select a historical point in time as the simulated "application date";

[0051] Collect all available data about the customer at the time of application and review the customer's actual performance within the time window. If the customer has reached the preset default threshold, the customer is deemed to have defaulted and a label "1" is generated. If the customer has consistently maintained normal repayments, the customer is deemed not to have defaulted and a label "0" is generated.

[0052] The debt default rate is obtained by statistically analyzing the total number of customers who defaulted and did not default within multiple time windows and comparing it with the number of defaults.

[0053] It should be noted that the credit assessment model library is constructed as follows:

[0054] B1: A sample of completed loan periods is extracted from the candidate customers and divided into the following categories in chronological order:

[0055] Training set (70%): Used for model parameter learning, selecting samples from earlier periods (such as data from 2021-2023).

[0056] Validation set (15%): used for hyperparameter tuning and model selection, selecting mid-term samples (such as data from 2023-2024).

[0057] Test set (15%): Used to evaluate the model's generalization ability, using the latest samples (such as data from 2024-2025).

[0058] It should be explained that dividing the effective risk control features of the candidate customers who have completed the loan cycle into training, validation, and test sets can ensure the generalization of the subsequent model construction, and dividing by time order can avoid time travel (the training set time must not be later than the validation / test set time), ensuring that the model can adapt to future risk changes.

[0059] B2: For each credit sub-model's risk dimension, specific input features are selected from effective risk control features. For example, the personal credit model focuses on an individual's repayment ability and willingness, selecting loan-to-value ratio and delinquency score as input features; the vehicle risk model reflects the risk of collateral value, selecting vehicle valuation volatility and vehicle depreciation rate as input features; the behavioral fraud model reflects a customer's personal behavioral default situation, selecting identity matching degree and debt default rate as input features. Then, the features are adapted, specifically by standardizing loan-to-value ratio, delinquency score, vehicle valuation volatility, vehicle depreciation rate, as well as identity matching degree and debt default rate, to ensure that features of different scales are treated fairly in model training, thereby accelerating model convergence and improving performance.

[0060] It should be further explained that the credit assessment model library contains three AI sub-models, each implementing assessment through differentiated algorithms, as detailed below:

[0061] The personal credit model uses gradient boosting decision trees, taking the loan-to-value ratio and delinquency score as inputs, thereby capturing the non-linear relationship between features. The optimal number of trees is determined by the validation set and iterated. The predicted values ​​of all trees are weighted and summed, and finally the self-default probability P(t) between 0 and 1 is output. Specifically, it is expressed as: P(t) = σ(Σ(predicted value of the i-th tree × weight i)), where σ is the Sigmoid function, which maps the summation result to the interval [0, 1].

[0062] For example, after training with a gradient boosting decision tree, the impact of loan-to-value ratio and delinquency score on the probability of default exhibits a significant non-linearity. When the delinquency score is ≤30: if the loan-to-value ratio is ≤70%, the probability of default is stable at 0.5%-1.2%, indicating low personal credit risk. If the loan-to-value ratio is >70%, the probability of default is 1.2%-3.8%. After the loan-to-value ratio exceeds 70%, the probability of default begins to rise, but the increase is gradual and still falls within the low-risk range. When the delinquency score is >70: if the loan-to-value ratio is ≤70%, the probability of default is 5.2%-6.8%; if the loan-to-value ratio is >70%, the probability of default is 12%-18%. At this point, high delinquency and high leverage in personal credit create a risk resonance. When the delinquency score is between 30 and 70: if the loan-to-value ratio is ≤70%, the probability of default is stable at 2.1%-4.5%, indicating low personal credit risk. If the loan-to-value ratio is >70%, the probability of default is 4.5%-15.2%.

[0063] When the loan-to-value ratio increases or decreases from 70%, the increase in the probability of default varies under different delinquency scoring scenarios. This phenomenon, where the same feature changes but has completely different impacts due to a different feature, is a typical example of interactive nonlinearity.

[0064] The vehicle risk model uses the random forest algorithm, taking vehicle valuation volatility and vehicle depreciation rate as input features, and the disposal loss rate as the objective. By minimizing the mean square error and splitting nodes, after multiple trees are trained independently, the model outputs the predicted disposal loss rate. The final prediction result is the average of the predicted disposal loss rate. The disposal loss rate c is obtained by subtracting the outstanding loan principal from the resale price after vehicle default and comparing it with the outstanding loan principal. The higher the loss rate, the greater the asset loss of the financial institution after vehicle default. The vehicle risk probability is then calculated using the formula P(c) = c * 0.8 + 0.2.

[0065] The behavioral fraud model uses the Isolation Forest algorithm, taking identity matching degree and debt default rate as input features, to construct multiple isolation trees. After the trees are trained, the path length *l* of each tree is extracted, the average path length is calculated, and then combined with the formula... Calculate the anomaly score S, where c(n) represents the average path length when the sample size is n, and collect the anomaly scores of all samples in the training set. Establish the mapping relationship between the anomaly score and the behavioral fraud probability, thereby obtaining the behavioral fraud probability.

[0066] For example, the mapping relationship between anomaly scores and behavioral fraud probabilities can be obtained through a mapping table, specifically represented as follows:

[0067] ,

[0068] Where P(q) represents the probability of fraudulent behavior. When S∈[0.8,1], it is mapped to a fraud risk probability of 85%. When S∈[0.5,0.8], it is mapped to a fraud risk probability of 30%. When S∈[0,0.5], it is mapped to a fraud risk probability of 2%.

[0069] It should be explained that gradient boosting trees, random forests, and isolated forests are all commonly used ensemble learning algorithms in machine learning. Gradient boosting trees are extremely effective at capturing nonlinear relationships and feature interactions. Random forests can be trained in parallel, are fast, reduce the risk of overfitting through multi-tree voting, and are insensitive to outliers and noisy data. Isolated forests do not require a large number of normal samples for training and have extremely high computational efficiency. The specific construction process of the model using each algorithm is existing technology and will not be elaborated here.

[0070] S3 enables professional and refined risk assessment across different dimensions by constructing a credit assessment model library containing multiple AI sub-models. The model library not only stores the trained sub-models but also supports real-time access, version management, and dynamic updates, providing standardized risk score outputs for subsequent comprehensive risk judgment.

[0071] S4: Comprehensive Judgment: Integrate the output results of the AI ​​sub-models, and use a non-linear evaluation model to aggregate the probability of default, vehicle risk probability and behavioral fraud probability for comprehensive judgment, thereby predicting the probability of default risk.

[0072] In this embodiment, the specific steps for predicting the probability of default risk need to be explained in detail as follows:

[0073] C1: The interaction terms are calculated based on the probability of default, the probability of vehicle risk, and the probability of behavioral fraud. Specifically, the interaction terms include: fraud-default P1 = P(q) × P(t), fraud-loss P2 = P(q) × P(c), and default-loss P3 = P(c) × P(t).

[0074] C2: Combine the original probability values ​​P(q), P(t), and P(c) with the calculated interaction terms P1, P2, and P3 to form a comprehensive feature vector X, specifically represented as: X = [P(q), P(t), P(c), P1, P2, P3], which captures the independent effects and interactions between personal credit, vehicle risk, and behavioral fraud.

[0075] C3: Input the comprehensive feature vector X into a pre-trained nonlinear evaluation model, which is usually a neural network with a nonlinear activation function. Its structure includes an input layer, a hidden layer and an output layer. The hidden layer uses an activation function to introduce a nonlinear transformation to capture complex patterns between features.

[0076] The nonlinear assessment model calculates the final default risk probability through forward propagation. Specifically, the model first calculates the hidden layer output h, expressed as: h = w(W1 × X + b1), where w is the activation function, and W1 and b1 represent the weights and biases of the hidden layer, respectively. Then, the output layer generates probabilities through the sigmoid function, expressed as:

[0077] ,

[0078] Where P represents the probability of default risk, W2 and b2 represent the weights and biases of the output layer, respectively, and e represents the natural constant. This probability P represents the overall risk of customer default after considering all factors.

[0079] It should be explained that the weights and biases of the hidden layers and the weights and biases of the output layers are trained using training methods known in the art, such as using the gradient descent algorithm to optimize the loss function and using the backpropagation algorithm to calculate and update the gradients of the weights (W1, W2) and biases (b1, b2) of each layer.

[0080] S5: Screening Creditworthy Customers: Maps the probability of default risk to a comprehensive credit score and automatically matches it with a preset risk level, thereby screening out creditworthy customers from the candidate customers.

[0081] In this embodiment, it should be specifically explained that the comprehensive credit score is obtained as follows:

[0082] Based on the probability of default risk, a comprehensive credit score is calculated using logarithmic odds transformation, specifically expressed as follows:

[0083] ,

[0084] Where G represents the comprehensive credit score, A and B represent the benchmark score and scaling factor, respectively. The benchmark score determines the center point of the score, the scaling factor determines the range of the score, and P represents the probability of default risk. Through logarithmic transformation, small changes in probability are amplified on the score, especially in extreme probability ranges.

[0085] It should be added that by adjusting the two parameters A and B, the score can be precisely controlled within the preset range, and the score distribution can be ensured to meet business expectations. For example, when P=0.5, G=A=50, when P decreases, S increases, and when P increases, S decreases.

[0086] It should be further explained that eligible customers are screened through a preset risk level. The preset risk level divides the scoring range into several risk levels based on business strategy and risk preference, as follows: If the comprehensive credit score is greater than or equal to 72, the used car financing risk is judged to be low; if the comprehensive credit score is between 50 and 72, the used car financing risk is judged to be medium; if the comprehensive credit score is less than 50, the used car financing risk is judged to be high. In this case, candidates with medium and low risk are selected as eligible customers.

[0087] In a specific example of the present invention, the lower the probability of default risk, the higher the comprehensive credit score. Assuming that A=50 and B=10 are adjusted according to the actual situation, P=0.1. Substituting into the comprehensive credit score calculation formula, the comprehensive credit score is G=50-10ln(0.11)=72, then the financial risk of used cars is low. If P=0.5, substituting into the comprehensive credit score calculation formula, the comprehensive credit score is G=50-10×0=50, then the financial risk of used cars is medium. If P=0.9, substituting into the comprehensive credit score calculation formula, the comprehensive credit score is G=50-10ln(9)=28, then the financial risk of used cars is high.

[0088] S6: Score Update: Regularly retrieve credit records, vehicle valuations, and vehicle condition history from third-party data sources for creditworthy users and update their comprehensive credit scores. Trigger risk warnings for creditworthy users whose identities change abnormally.

[0089] In this embodiment, it should be specifically explained that the abnormal change of identity of the creditworthy user is determined by periodically calculating its secondary comprehensive credit score, and then subtracting the secondary comprehensive credit score from the comprehensive credit score to obtain the credit comparison difference. If the credit comparison difference is 0 or positive, it indicates that the used car financial risk corresponding to the creditworthy user is stable or reduced. If the credit comparison difference is negative, it indicates that the used car financial risk corresponding to the creditworthy user is increased. At this time, it is determined that the creditworthy user has an abnormal change of identity.

[0090] As attached Figure 2 The AI-based used car financing credit scoring system shown also includes:

[0091] Data retrieval module: With user authorization, an access point is set up on the credit application page of the financial platform, allowing applicants to submit basic personal information and financial information, and authorizing the platform to retrieve credit records, vehicle valuations and vehicle condition history from third-party data sources;

[0092] Preliminary screening module: Based on preset data processing rules, it retrieves credit records, vehicle valuations, and vehicle condition history from third-party data sources to identify credit scenarios, and then preliminarily screens out applicants who have basic credit qualifications and marks them as candidate customers;

[0093] Model building module: retrieves the corresponding effective risk control features of candidate customers, and then builds a credit assessment model library. The credit assessment model library is used to store and run AI sub-models, including personal credit model, vehicle risk model and behavioral fraud model.

[0094] Comprehensive judgment module: It integrates the output results of the AI ​​sub-models and uses a non-linear evaluation model to aggregate its own default probability, vehicle risk probability, and behavioral fraud probability for comprehensive judgment, thereby predicting the default risk probability;

[0095] Secondary screening module: Maps the probability of default risk to a comprehensive credit score and automatically matches it with a preset risk level, thereby selecting creditworthy customers from the candidate customers;

[0096] Credit score update module: Regularly retrieves credit records, vehicle valuations, and vehicle condition history from third-party data sources for creditworthy users and updates their comprehensive credit score. Triggers risk warnings for creditworthy users whose identities change abnormally.

[0097] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other.

[0098] In conclusion, 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, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An AI-based used car financial credit scoring method, characterized in that, include: S1: Data Retrieval: With user authorization, an access point is set up on the credit application page of the financial platform, allowing applicants to submit basic personal information and financial information, and authorizing the platform to retrieve credit records, vehicle valuations, and vehicle condition history from third-party data sources; S2: Preliminary screening: Based on preset data processing rules, credit records, vehicle valuations, and vehicle condition history are retrieved from third-party data sources to identify credit scenarios, thereby initially screening applicants who have basic credit qualifications and marking them as candidate customers; S3: Model Building: Retrieve the corresponding effective risk control features of the candidate customers, and then build a credit assessment model library. The credit assessment model library is used to store and run AI sub-models, including personal credit model, vehicle risk model and behavioral fraud model. S4: Comprehensive Judgment: Integrate the output results of the AI ​​sub-models, and use a non-linear evaluation model to aggregate the probability of default, vehicle risk probability and behavioral fraud probability for comprehensive judgment, thereby predicting the probability of default risk. S5: Screening Creditworthy Customers: Mapping the probability of default risk to a comprehensive credit score and automatically matching it with a preset risk level, thereby screening out creditworthy customers from the candidate customers; S6: Score Update: Regularly retrieve credit records, vehicle valuations, and vehicle condition history from third-party data sources for creditworthy users and update their comprehensive credit scores. Trigger risk warnings for creditworthy users whose identities change abnormally.

2. The AI-based used car financial credit scoring method according to claim 1, characterized in that: The specific steps for identifying the credit scenario are as follows: A1: Categorize and analyze credit records, vehicle valuations, and vehicle condition history, and identify key risk signals from the characteristics of the categorization; A2: Count the number of key risk signals identified, and match each risk signal with the typical risk signals of different credit scenarios in the scenario rule base to obtain the credit scenario matched by each risk signal; A3: Analyze the frequency of occurrence of each credit scenario that matches each risk signal, and then take the scenario with the highest frequency of occurrence as the applicant's main credit scenario.

3. The AI-based used car financial credit scoring method according to claim 1, characterized in that: If the applicant's primary credit scenario is positive, the applicant is selected and marked as a candidate customer. If the applicant's primary credit scenario is negative, the applicant's used car sales are automatically rejected.

4. The AI-based used car financial credit scoring method according to claim 1, characterized in that: The credit assessment model library is constructed as follows: B1: A sample of completed loan periods is extracted from the candidate customers and divided into the following categories in chronological order: Training set: Used for learning model parameters, it selects samples from earlier periods; Validation set: used for hyperparameter tuning and model selection, selecting samples from the intermediate period; Test set: Used to evaluate the model's generalization ability, selecting the latest samples; B2: For each credit sub-model's risk dimension, select exclusive input features from effective risk control features.

5. The AI-based used car financial credit scoring method according to claim 4, characterized in that: The credit assessment model library contains three AI sub-models, each employing a differentiated algorithm for assessment, as detailed below: The personal credit model uses gradient boosting decision trees, taking the loan-to-value ratio and delinquency score as inputs, thereby capturing the non-linear relationship between features. The optimal number of trees is determined by the validation set and iterated. The predicted values ​​of all trees are weighted and summed, and finally the default probability P(t) between 0 and 1 is output. The vehicle risk model uses the random forest algorithm, taking vehicle valuation volatility and vehicle depreciation rate as input features, with disposal loss rate as the objective. By minimizing the mean square error to split nodes, the final prediction result is the average value of the predicted disposal loss rate. Then, the vehicle risk probability is calculated using the formula P(c) = c*0.8 + 0.

2. The behavioral fraud model uses the isolated forest algorithm to construct multiple isolated trees, taking identity matching degree and debt default rate as input features. After the trees are trained, the path length l of each tree is extracted, the anomaly score S is calculated, and the anomaly scores of all samples in the training set are collected to establish a mapping relationship between the anomaly score and the behavioral fraud probability, thereby obtaining the behavioral fraud probability.

6. The AI-based used car financial credit scoring method according to claim 1, characterized in that: The specific steps for predicting the probability of default risk are as follows: C1: The interaction terms are calculated based on the probability of default, the probability of vehicle risk, and the probability of behavioral fraud. Specifically, the interaction terms include: fraud-default P1 = P(q) × P(t), fraud-loss P2 = P(q) × P(c), and default-loss P3 = P(c) × P(t). C2: Combine the original probability values ​​P(q), P(t), and P(c) with the calculated interaction terms P1, P2, and P3 to form a comprehensive feature vector X, specifically represented as: X = [P(q), P(t), P(c), P1, P2, P3]; C3: The comprehensive feature vector X is input into a pre-trained nonlinear evaluation model. The nonlinear evaluation model calculates the final default risk probability through forward propagation. Specifically, the model first calculates the hidden layer output h, expressed as: h = w(W1 × X + b1), where w is the activation function, and W1 and b1 represent the weights and biases of the hidden layer, respectively. Then, the output layer generates probabilities through the sigmoid function, expressed as: ， Where P represents the probability of default risk, W2 and b2 represent the weights and biases of the output layer, respectively, and e represents the natural constant.

7. The AI-based used car financial credit scoring method according to claim 1, characterized in that: The comprehensive credit score is obtained in the following ways: Based on the probability of default risk, a logarithmic odds transformation is used to amplify small changes in probability and calculate a comprehensive credit score.

8. The AI-based used car financial credit scoring method according to claim 1, characterized in that: The creditworthy customers are screened through a preset risk level, and the scoring range is divided into several risk levels, as follows: if the comprehensive credit score is greater than or equal to 72, the used car finance risk is judged to be low; if the comprehensive credit score is between 50 and 72, the used car finance risk is judged to be medium; if the comprehensive credit score is less than 50, the used car finance risk is judged to be high. At this time, the candidates with medium and low risk are selected as creditworthy customers.

9. A used car financial credit scoring system based on any one of claims 1 to 8, characterized in that: Data retrieval module: With user authorization, an access point is set up on the credit application page of the financial platform, allowing applicants to submit basic personal information and financial information, and authorizing the platform to retrieve credit records, vehicle valuations and vehicle condition history from third-party data sources; Preliminary screening module: Based on preset data processing rules, it retrieves credit records, vehicle valuations, and vehicle condition history from third-party data sources to identify credit scenarios, and then preliminarily screens out applicants who have basic credit qualifications and marks them as candidate customers; Model building module: retrieves the corresponding effective risk control features of candidate customers, and then builds a credit assessment model library. The credit assessment model library is used to store and run AI sub-models, including personal credit model, vehicle risk model and behavioral fraud model. Comprehensive judgment module: It integrates the output results of the AI ​​sub-models and uses a non-linear evaluation model to aggregate its own default probability, vehicle risk probability, and behavioral fraud probability for comprehensive judgment, thereby predicting the default risk probability; Secondary screening module: Maps the probability of default risk to a comprehensive credit score and automatically matches it with a preset risk level, thereby selecting creditworthy customers from the candidate customers; Credit score update module: Regularly retrieves credit records, vehicle valuations, and vehicle condition history from third-party data sources for creditworthy users and updates their comprehensive credit score. Triggers risk warnings for creditworthy users whose identities change abnormally.

Images