C card strategy method based on marginal revenue algorithm

Abstract

The invention discloses a C card strategy method based on a marginal income algorithm, and relates to the technical field of credit product business management, and the method comprises the steps: obtaining original data of a customer, carrying out the preprocessing, and generating a standardized data set; based on the standardized data set, respectively constructing a marginal income prediction model and a credit risk prediction model; utilizing a marginal income prediction model and a credit risk prediction model to predict marginal income and default probability of credit product customers; according to the marginal income and the default probability, a differential credit granting strategy of the credit product customers is formulated; and verifying the validity of the differentiated credit granting strategy through strategy simulation or historical data, and performing optimization adjustment according to a verification result. According to the method, the accuracy of marginal income prediction is improved through the double-layer feature selection mechanism and the XGBoost algorithm in combination with the design of the income correction sub-model, and the potential value of the customer can be identified and evaluated more accurately.

Description

Technical Field

[0001] The invention relates to the technical field of credit product business management, and in particular to a C card strategy method based on a marginal revenue algorithm. Background Art

[0002] In recent years, with the rapid development of credit product business, the scale of credit and transaction volume have shown explosive growth, and the bank's credit business center is facing the dual pressure of business expansion and risk control. In this context, formulating a scientific credit business strategy is of great significance to improving the bank's operating efficiency and controlling risks.

[0003] At present, the credit business strategy methods commonly used in the industry have the following shortcomings: business decisions are too mechanical and fail to fully consider the actual transaction behavior and risk characteristics of customers; traditional scoring card models mainly rely on historical data and have limited predictive capabilities for the future; there is a lack of quantitative correlation between business strategies and bank profit targets; risk control measures are single and a dynamic risk warning mechanism has not been established; there is a lack of accurate calculation and optimization methods for marginal benefits, resulting in inefficient resource allocation. Summary of the invention

[0004] In view of the problems that the existing credit product business strategy method has in mechanical decision-making, limited prediction ability, single risk control, etc., the present invention is proposed.

[0005] Therefore, the problem to be solved by the present invention is how to formulate differentiated credit strategies based on customer data from the two dimensions of marginal revenue and credit risk.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In the first aspect, an embodiment of the present invention provides a C-card strategy method based on a marginal benefit algorithm, which includes obtaining and preprocessing original customer data to generate a standardized data set, wherein the original customer data includes customer attribute data, transaction flow data, credit assessment data, and historical default data; based on the standardized data set, constructing a marginal benefit prediction model and a credit risk prediction model respectively; using the marginal benefit prediction model and the credit risk prediction model to predict the marginal benefit and default probability of credit product customers; formulating differentiated credit strategies for credit product customers based on marginal benefits and default probabilities; verifying the effectiveness of the differentiated credit strategy through strategy simulation or historical data, and optimizing and adjusting according to the verification results.

[0008] As a preferred solution of the C-card strategy method based on the marginal benefit algorithm described in the present invention, the construction process of the marginal benefit prediction model is as follows: using the customer attribute data and transaction flow data in the standardized data set, a two-layer feature selection mechanism is constructed to generate a first feature subset; according to the transaction flow data in the standardized data set, the marginal contribution value of each customer's credit product is calculated to generate customer marginal benefit data; an XGBoost prediction base model is constructed, and the first feature subset and customer marginal benefit data are used for training; using the credit assessment data in the standardized data set, a benefit correction submodel is constructed to calculate the benefit adjustment factor; the XGBoost prediction base model and the benefit correction submodel are integrated to obtain a marginal benefit prediction model.

[0009] As a preferred solution of the C-card strategy method based on the marginal benefit algorithm described in the present invention, the construction process of the XGBoost prediction base model is as follows: constructing a first training feature matrix based on the first feature subset, and taking the customer marginal benefit data as the first training label; dividing the first training feature matrix and its corresponding first training label into a first training subset and a first validation subset according to a preset ratio; constructing a first decision tree, and using the first training subset to calculate the splitting gain of the first decision tree, and determining the optimal splitting feature and splitting threshold; iteratively constructing a decision tree sequence based on the residual of the previous decision tree until a preset number of iterations is reached; combining the decision tree sequence to form the XGBoost prediction base model, and outputting the initial marginal benefit prediction value.

[0010] As a preferred solution of the C-card strategy method based on the marginal benefit algorithm described in the present invention, the construction process of the credit risk prediction model is as follows: construct a customer transaction association network using transaction flow data in a standardized data set; use a community discovery algorithm to divide the customer transaction association network into groups and generate group identifiers; extract customer node association features based on the customer transaction association network, and generate group features in combination with the group identifiers; calculate risk transmission features based on the customer transaction association network; determine whether the customer has defaulted based on the transaction flow data in the standardized data set, and generate a default identifier; construct a second training feature matrix based on customer node association features, group features and risk transmission features, and divide the second training feature matrix into a second training subset and a second verification subset according to a preset ratio; construct and train a credit risk prediction model based on a gradient boosting framework.

[0011] As a preferred solution of the C-card strategy method based on the marginal benefit algorithm described in the present invention, wherein: constructing and training a credit risk prediction model based on a gradient boosting framework includes the following steps: determining the initial default probability prediction value p0 of all samples; calculating the negative gradient value of each sample in the second training subset based on the log-likelihood loss function; using the negative gradient value as the target variable, combining the customer node association characteristics, group characteristics and risk conduction characteristics to construct a CART decision tree, obtain a risk prediction base learner, and output the t-th round prediction adjustment value Δp t ; The product of the output result of the risk prediction base learner and the preset learning rate is used as the default probability adjustment amount for the current round; the default probability adjustment amount is added to the default probability prediction value of the previous round to obtain the updated default probability prediction value; AUC is used as the model evaluation indicator, and the negative gradient calculation and model iterative update process are repeated until the convergence conditions are met to obtain the final credit risk prediction model.

[0012] As a preferred solution of the C-card strategy method based on the marginal benefit algorithm described in the present invention, the formulation of differentiated credit strategies for credit product customers includes the following steps: determining the marginal benefit group to which the customer belongs based on the marginal benefit forecast value, the marginal benefit group includes a high marginal benefit group, a medium marginal benefit group and a low marginal benefit group; formulating a basic credit strategy based on the marginal benefit group; adjusting the basic credit strategy in combination with the probability of default to form a differentiated credit strategy; and outputting a differentiated credit strategy report.

[0013] As a preferred solution of the C-card strategy method based on the marginal benefit algorithm described in the present invention, the convergence condition includes that the number of iterations of the risk prediction base learner reaches a preset number of iterations, or the change value of the AUC indicator on the second verification subset of the risk prediction base learner in consecutive preset rounds of iterations is less than a preset convergence threshold.

[0014] In the second aspect, an embodiment of the present invention provides a C-card policy system based on a marginal benefit algorithm, which includes a data processing module for acquiring and preprocessing original customer data to generate a standardized data set; a model building module for constructing a marginal benefit prediction model and a credit risk prediction model based on the standardized data set, respectively; a model prediction module for predicting the marginal benefit and default probability of credit product customers using the marginal benefit prediction model and the credit risk prediction model; a strategy formulation module for formulating differentiated credit strategies for credit product customers based on marginal benefits and default probabilities; and a verification module for verifying the effectiveness of differentiated credit strategies through strategy simulation or historical data, and optimizing and adjusting according to the verification results.

[0015] The beneficial effects of the present invention are as follows: the present invention improves the accuracy of marginal benefit prediction through a double-layer feature selection mechanism and an XGBoost algorithm, combined with the design of a benefit correction sub-model, and can more accurately identify and evaluate the potential value of customers. The social network analysis method is innovatively introduced. By constructing a customer transaction association network and extracting group characteristics and risk transmission characteristics, the impact of the correlation between customers on credit risk is deeply explored, and the comprehensiveness of risk prediction is improved. Combining the dual dimensions of marginal benefit and credit risk, a hierarchical strategy system based on beneficiary groups and risk levels is designed, achieving more refined customer differentiation management. It has strong practicality and operability. Through strategy simulation and historical data verification mechanisms, the credit strategy can be continuously optimized and adjusted to ensure the effectiveness and adaptability of strategy execution. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0017] Figure 1 This is a framework flow chart of the C-card strategy method based on the marginal benefit algorithm.

[0018] Figure 2 This is a flow chart for calculating the customer marginal revenue forecast value of the C-card strategy method based on the marginal revenue algorithm.

[0019] Figure 3 The flowchart for constructing the credit risk prediction model of the C-card strategy method based on the marginal benefit algorithm. DETAILED DESCRIPTION

[0020] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0021] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0022] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0023] Example 1, reference Figure 1 to Figure 3 , which is the first embodiment of the present invention, and provides a C-card strategy method based on a marginal benefit algorithm. The framework flow chart is as follows Figure 1 Shown include,

[0024] S1: Obtain customer raw data and preprocess it to generate a standardized data set for strategy analysis.

[0025] Among them, the original customer data includes customer attribute data, transaction flow data, credit assessment data, and historical default data. The preprocessing work includes data quality inspection, data cleaning, and data integration.

[0026] For example, customer attribute data includes but is not limited to customer age, gender, income level, education level, and occupation type; transaction flow data includes but is not limited to transaction amount, transaction time, transaction type, merchant type, transaction frequency, and repayment record; credit assessment data includes but is not limited to credit score, number of overdue payments, maximum number of overdue days, credit product quota utilization rate, and other loan conditions; historical default data includes but is not limited to the number of historical defaults, default amount, default duration, and default relief method.

[0027] S2: Based on the standardized data set, construct marginal profit prediction model and credit risk prediction model respectively.

[0028] Specifically, the following steps are included:

[0029] S2.1: Use machine learning algorithms to build a marginal revenue prediction model based on a standardized data set.

[0030] Specifically, the calculation process of the customer marginal revenue forecast value is as follows: Figure 2 As shown, the following steps are included:

[0031] S2.1.1: Using the customer attribute data and transaction flow data in the standardized data set, a two-layer feature selection mechanism is constructed to generate the first feature subset.

[0032] Specifically, the variance selection method is first used to remove low-discrimination features with variance less than 0.01, where the variance threshold of 0.01 is the optimal threshold determined by feature distribution analysis; then LASSO regression is used for feature screening, and features with coefficients not 0 are selected. The regularization parameter of LASSO is determined by cross-validation; finally, XGBoost's feature importance score is used to select features with importance scores greater than 0.05, and finally the first feature subset is obtained. At the same time, correlation analysis is performed on the selected features, and highly correlated features with correlation coefficients greater than 0.8 are eliminated.

[0033] In this embodiment, the first feature subset includes, but is not limited to, average monthly transaction amount, transaction frequency, repayment timeliness rate, credit product quota utilization rate, customer annual income, and historical default times.

[0034] S2.1.2: Based on the transaction flow data in the standardized data set, calculate the marginal contribution value of each customer's credit product and generate customer marginal profit data.

[0035] Specifically, the calculation formula for the marginal contribution value of a credit product is as follows:

[0036] M value =(M af +M ai +M ac )×(1-Loss gd )

[0037] Among them, M value is the marginal contribution value of the credit product, M af is the annualized fee income, M ai is the annualized interest income, M ac is the annualized installment income, Loss gd is the default loss rate, and the specific formula is as follows:

[0038] Loss gd =T hdl / T hcg

[0039] Among them, T hdl is the total historical default loss (in the last 12 months), T hcg It is the total amount of historical credit (last 12 months).

[0040] S2.1.3: Build an XGBoost prediction base model and use the first feature subset and customer marginal revenue data for training.

[0041] Preferably, the construction process of the XGBoost prediction base model is as follows: construct a first training feature matrix based on the first feature subset, and use the customer marginal revenue data as the first training label; divide the first training feature matrix and its corresponding first training label into a first training subset and a first validation subset according to a preset ratio, and in this embodiment, the preset ratio is set to 8:2; construct a first decision tree, use the first training subset to calculate the split gain of the first decision tree, and determine the optimal split feature and split threshold. The specific calculation formula is as follows:

[0042]

[0043] Among them, GL and GR are the sum of the first-order derivatives of the left child node and the right child node respectively, HL and HR are the sum of the second-order derivatives of the left child node and the right child node respectively, λ is the L2 regularization term coefficient, γ is the minimum gain threshold for node splitting, and Gain is the splitting gain.

[0044] Furthermore, a decision tree sequence is constructed based on the residual iteration of the previous decision tree until a preset number of iterations is reached, and the preferred preset number of iterations is 100; the decision tree sequence is combined to form an XGBoost prediction base model, and the initial marginal benefit prediction value is output.

[0045] Moreover, the objective function of the XGBoost prediction base model consists of a loss function and a regularization term, as follows:

[0046]

[0047] Among them, y i is the actual marginal benefit value, To predict the marginal revenue value, w j is the weight of the jth leaf node, and T is the number of leaf nodes. The objective function consists of two parts: the first term is the mean square loss function, which is used to measure the difference between the model prediction value and the true value; the second term is the regularization term, where λw j 2 Overfitting is prevented by penalizing the square of the leaf node weights, and γT controls the complexity of the tree by penalizing the number of leaf nodes.

[0048] During the training process, the model parameters were set as follows: learning rate was 0.1, maximum tree depth was 6, minimum split gain threshold was 0.1, L2 regularization coefficient was 1.0, node split minimum gain threshold was 0.3, subsampling ratio for each iteration was 0.8, and feature sampling ratio was 0.8. In addition, an early stopping strategy was adopted, and training was stopped when the evaluation indicators on the validation subset did not improve for 5 consecutive rounds. The evaluation indicators included root mean square error, absolute error of evaluation, and determination coefficient.

[0049] Preferably, the present invention can effectively capture the nonlinear relationship between customer characteristics and marginal revenue and improve prediction accuracy by constructing an XGBoost prediction base model. At the same time, through reasonable regularization design and early stopping strategy, it can effectively prevent model overfitting and ensure the generalization ability of the model.

[0050] S2.1.4: Use the credit assessment data in the standardized data set to construct a return correction sub-model and calculate the return adjustment factor.

[0051] Specifically, the construction process of the income correction sub-model is as follows: a linear regression model is constructed based on the credit assessment data, with the credit score as the independent variable and the ratio of actual income to predicted income as the dependent variable, and the income adjustment factor is calculated. The specific formula is as follows:

[0052] α=β0+β1×S c +β2×N d +β3×D max

[0053] Among them, β0, β1, β2, β3 are linear regression coefficients, S c is the credit score, N d is the number of overdue payments, D max is the maximum number of overdue days, and α is the yield adjustment factor.

[0054] S2.1.5: Integrate the XGBoost prediction base model and the revenue correction sub-model to obtain the marginal revenue prediction model.

[0055] It should be noted that the integration method is: based on the income adjustment factor of the income correction sub-model, the output of the XGBoost prediction base model is corrected. The formula is as follows:

[0056] F=α·P

[0057] Among them, P is the predicted value output by the XGBoost prediction base model, α is the profit adjustment factor, and F is the predicted value of customer marginal profit.

[0058] S2.2: Use social network analysis and group clustering methods to build a credit risk prediction model based on a standardized data set.

[0059] Specifically, the construction flow chart of the credit risk prediction model is as follows: Figure 3 As shown, the following steps are included:

[0060] S2.2.1: Use transaction flow data in the standardized data set to construct a customer transaction association network.

[0061] Among them, the nodes of the customer transaction network are customers, and the edges are transaction relationships, including transaction amount, transaction frequency and transaction time interval.

[0062] Specifically, the construction process of the customer transaction association network is as follows: first, each transaction record is converted into a network edge, in which the customer ID and merchant ID are used as the source node and the target node respectively; then, for node pairs with multiple transactions, the cumulative transaction amount is calculated as the edge weight, the cumulative number of transactions is used as the transaction frequency attribute of the edge, and the average time interval between adjacent transactions is used as the time attribute of the edge; finally, a weighted undirected graph is constructed based on the above edge attributes as the customer transaction association network.

[0063] S2.2.2: Use the Fast Unfolding community discovery algorithm to divide the customer transaction association network into groups and generate group identifiers.

[0064] Specifically, the execution process of the Fast Unfolding algorithm is as follows: first, a unique community identifier is assigned to each node; then two phases are iterated: in the first phase, all nodes are traversed, the modularity gain of moving the node to the adjacent community is calculated, and the moving plan with the largest gain is selected; in the second phase, a new network is constructed based on the results of the first phase, and nodes belonging to the same community are merged into a new node. These two phases are repeated until the modularity gain is no longer significantly improved (in this embodiment, an incremental threshold is set to determine the improvement range) or the preset number of iterations is reached (set to 100 times in this embodiment).

[0065] In addition, the calculation formula of the modularity gain ΔQ is as follows:

[0066]

[0067] Among them, ∑ in is the weight sum of the edges within the community, ∑ tot is the sum of the weights of all edges connected to the community, k i,in is the weight sum of the edges connecting node i and the target community node, k i is the sum of the edge weights of node i, and m is the sum of all edge weights in the network.

[0068] S2.2.3: Extract customer node association features based on the customer transaction association network, and generate group features in combination with group identification.

[0069] Among them, customer node association characteristics include degree centrality, betweenness centrality, closeness centrality and clustering coefficient, and group characteristics include group size, group density, group default rate and group average transaction amount.

[0070] S2.2.4: Based on the customer transaction association network, use the PageRank algorithm to calculate the risk transmission characteristics.

[0071] Specifically, based on the customer transaction association network G(V,E), the risk transmission characteristic value of each customer node is calculated. For each node v∈V, its risk score is initialized to r0(v), the initial risk score of the defaulted customer is 1, and the initial risk score of the non-defaulted customer is 0. Using the damping coefficient θ=0.85, the risk score of each node is updated through iterative calculation:

[0072]

[0073] Among them, r t+1(v) is the risk score of node v at iteration t+1, (1-θ)r0(v) is the contribution of the initial risk score, θ is the damping coefficient, ∑ (u,v)∈E To traverse all edges pointing to node v, ω(u,v) represents the weight of edge (u,v), that is, the cumulative transaction amount, r t (u) is the risk score of node u at the tth iteration, ∑ (u,x)∈E ω(u,x) is the sum of the weights of all outgoing edges of node u, which is used for normalization.

[0074] Furthermore, the iteration is stopped when the update amplitude of the risk score is less than a preset threshold (set to 1e-6 in this embodiment) or when the maximum number of iterations (set to 100 in this embodiment) is reached. Finally, the risk score r(v) of each node is the risk transmission characteristic value of the customer.

[0075] S2.2.5: Based on the transaction flow data in the standardized data set, identify whether the customer has defaulted and generate a default mark.

[0076] It should be noted that the default flag includes 0 and 1, where 0 indicates no default and 1 indicates default.

[0077] Moreover, a default will be determined if any of the following rules is met: a single payment is overdue for more than 90 days, three consecutive repayments are overdue for more than 30 days, the cumulative overdue amount exceeds 50% of the credit limit, or there is a record of guarantee compensation.

[0078] S2.2.6: Based on the customer node association characteristics, group characteristics and risk transmission characteristics, a second training feature matrix is ​​constructed, and the second training feature matrix is ​​divided into a second training subset and a second verification subset according to a preset ratio.

[0079] Specifically, the construction process of the second training feature matrix is ​​as follows: customer node association features (4), group features (4) and risk conduction features (1) are combined to form a 9-dimensional feature vector, where each row represents a customer sample, and standardized processing is performed to obtain the final second training feature matrix.

[0080] S2.2.7: Build and train a credit risk prediction model based on the gradient boosting framework.

[0081] Preferably, the ratio of the number of samples with a default flag of 1 to the total number of samples in the second training subset is calculated as the initial default probability prediction value p0 of all samples, and it is verified whether p0 is in the interval [0,1]; the negative gradient value of each sample in the second training subset is calculated based on the log-likelihood loss function; the negative gradient value is used as the target variable, and a CART decision tree is constructed in combination with customer node association characteristics, group characteristics and risk conduction characteristics to obtain a risk prediction base learner, and the prediction adjustment value Δp for the tth round is output. t; The product of the output result of the risk prediction base learner and the preset learning rate (value range 0-1) is used as the default probability adjustment amount for the current round; the default probability adjustment amount is added to the default probability prediction value of the previous round to obtain the updated default probability prediction value; AUC is used as the model evaluation indicator, and the negative gradient calculation and model iterative update process are repeated until the convergence conditions are met to obtain the final credit risk prediction model.

[0082] Among them, the convergence conditions include that the number of iterations of the risk prediction base learner reaches a preset number of iterations, or that the change value of the AUC indicator on the second verification subset of the risk prediction base learner in consecutive preset rounds of iterations is less than a preset convergence threshold.

[0083] It should be noted that the process of obtaining the risk prediction base learner is as follows: the negative gradient value is used as the target variable, and the customer node association features, group features and risk conduction features are used as input variables to form a training data set. Initialize the root node based on the training data set, which contains all training samples. For the current node, traverse all input features and their possible splitting points, calculate the Gini index corresponding to each feature splitting point, select the feature and splitting point with the smallest Gini index as the optimal splitting scheme, and divide the samples into left and right child nodes according to the optimal splitting scheme. Repeat the process of feature selection and node splitting until the stopping condition is met, including reaching the maximum depth, the number of node samples is less than the minimum sample number threshold, or the optimal splitting gain of all features is less than the threshold. For all leaf nodes, calculate the average negative gradient value of the samples on the node as the predicted value of the node. The complete CART decision tree is obtained as the risk prediction base learner, and the predicted value of each sample is its prediction adjustment value Δp for the tth round. t .

[0084] In addition, since positive and negative samples (default and non-default samples) in credit risk prediction are usually unevenly distributed, AUC is not affected by the sample distribution and can comprehensively evaluate the performance of the model under different thresholds, making it more suitable as an evaluation indicator for risk prediction models.

[0085] Preferably, the present invention can achieve accurate quantitative assessment of customer default risk through a credit risk prediction model, effectively integrate customer node association characteristics, group characteristics and risk transmission characteristics, improve model prediction capabilities through iterative optimization, and combine marginal benefit prediction results to provide a reliable quantitative basis for customer strategy optimization and risk management decision-making.

[0086] S3: Use marginal revenue prediction models and credit risk prediction models to predict the marginal revenue and default probability of credit product customers.

[0087] S4: Develop differentiated credit granting strategies for credit product customers based on marginal returns and default probabilities.

[0088] Specifically, the following steps are included:

[0089] S4.1: Based on the marginal benefit forecast, determine the marginal beneficiary group to which the customer belongs.

[0090] Among them, the marginal beneficiary groups include high marginal beneficiary groups, medium marginal beneficiary groups and low marginal beneficiary groups.

[0091] It should be noted that the division of marginal beneficiary groups is determined by the quantiles of the marginal benefit prediction values, as follows: after sorting the marginal benefit prediction values, the first 30% are taken as the high marginal benefit group, the bottom 30% are taken as the low marginal benefit group, and the rest are the medium marginal benefit groups.

[0092] S4.2: Develop basic credit strategies based on marginal beneficiary groups.

[0093] Specifically, for the high marginal beneficiary groups, incentive measures such as increasing credit limits, reducing annual fees, and lowering installment rates will be adopted; for the medium marginal beneficiary groups, the existing credit limits and rate policies will be maintained; for the low marginal beneficiary groups, the credit limits will be appropriately reduced, and the annual fees and installment rates will be increased.

[0094] S4.3: Based on the probability of default, adjust the basic credit strategy to form a differentiated credit strategy.

[0095] Specifically, the probability of default is divided into three levels: low risk, medium risk and high risk according to the two thresholds of 0.3 and 0.7; for low-risk customers, the basic credit strategy of the corresponding marginal beneficiary group is implemented; for medium-risk customers, the credit limit is reduced by 20% and the installment rate is increased by 30% on the basis of the basic credit strategy; for high-risk customers, the credit limit is reduced by 50% and the installment rate is increased by 50% on the basis of the basic credit strategy, and transaction restrictions are increased, such as restrictions on cash withdrawals and restrictions on large transactions.

[0096] S4.4: Output differentiated credit strategy report.

[0097] Among them, the differentiated credit strategy report includes basic customer information, marginal profit forecast value, default probability forecast value, beneficiary group category, risk level, adjusted credit limit, annual fee policy, installment rate, restrictive measures, and strategy adjustment suggestions.

[0098] S5: Verify the effectiveness of differentiated credit strategies through strategy simulation or historical data, and make optimization adjustments based on the verification results.

[0099] Specifically, the strategy verification process first randomly extracts 3 months of credit records from historical data, and divides the experimental group and the control group in a ratio of 7:3 to ensure that the two groups have similar distributions in terms of credit limit, industry, region, etc. The experimental group adopts a new differentiated credit strategy, and the control group maintains the original strategy. Through a 90-day comparative experiment, key indicators such as credit limit utilization rate, overdue rate, and profit margin are monitored weekly. After the experiment, a statistical significance test is conducted to evaluate the strategy effect, and the parameters and rules are optimized based on the experimental data to form the next round of iterative plans.

[0100] Furthermore, this embodiment also provides a C-card policy system based on the marginal benefit algorithm, including a data processing module, which is used to obtain customer original data and perform preprocessing to generate a standardized data set; a model building module, which is used to construct a marginal benefit prediction model and a credit risk prediction model based on the standardized data set; a model prediction module, which is used to use the marginal benefit prediction model and the credit risk prediction model to predict the marginal benefit and default probability of credit product customers; a strategy formulation module, which is used to formulate differentiated credit strategies for credit product customers based on marginal benefits and default probabilities; a verification module, which is used to verify the effectiveness of differentiated credit strategies through strategy simulation or historical data, and to optimize and adjust according to the verification results.

[0101] In summary, the present invention improves the accuracy of marginal benefit prediction through a double-layer feature selection mechanism and XGBoost algorithm, combined with the design of a benefit correction sub-model, and can more accurately identify and evaluate the potential value of customers. The social network analysis method is innovatively introduced. By constructing a customer transaction association network and extracting group characteristics and risk transmission characteristics, the impact of the correlation between customers on credit risk is deeply explored, and the comprehensiveness of risk prediction is improved. Combining the dual dimensions of marginal benefit and credit risk, a hierarchical strategy system based on beneficiary groups and risk levels is designed to achieve more refined customer differentiation management. It has strong practicality and operability. Through strategy simulation and historical data verification mechanism, the credit strategy can be continuously optimized and adjusted to ensure the effectiveness and adaptability of strategy execution.

[0102] Example 2, reference Figure 1 to Figure 3 , which is the second embodiment of the present invention, and this embodiment provides a C-card strategy method based on a marginal benefit algorithm. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0103] The experiment used real credit product business data from a large commercial bank. The dataset contains customer attributes, transaction information, credit records, etc. of 100,000 credit product holders, ranging from January 1, 2019 to December 31, 2023. By preprocessing the dataset, a standardized dataset for analysis was generated.

[0104] On this basis, the present invention first uses a machine learning algorithm to build a marginal revenue prediction model. Specifically, the present invention uses the XGBoost model to model the relationship between customer characteristics and marginal revenue, and introduces credit scores as adjustment factors to further improve the accuracy of the prediction. The marginal revenue prediction model achieves a mean absolute error of less than 5% on the validation set.

[0105] Secondly, the present invention constructs a credit risk prediction model based on the customer transaction network analysis and group clustering method. The model comprehensively considers customer association characteristics, group characteristics and risk transmission characteristics, and uses the gradient boosting algorithm for training and iterative optimization. The AUC index on the validation set reaches above 0.85.

[0106] By integrating the prediction results of the above two models, the present invention designs a differentiated credit product credit strategy. Specifically, the present invention first divides customers into three beneficiary groups: high, medium and low according to the marginal benefit prediction value. Differentiated credit strategies are adopted for different beneficiary groups: for high marginal beneficiary groups, preferential measures such as increasing credit limits, reducing annual fees, and reducing installment rates are adopted; for medium marginal beneficiary groups, the existing credit policy is maintained; for low marginal beneficiary groups, the credit limits are appropriately reduced, and annual fees and installment rates are increased.

[0107] On this basis, the present invention further adjusts the above basic credit strategy in combination with the customer's default risk level. For low-risk customers, the basic strategy of the corresponding beneficiary group is implemented; for medium-risk customers, the credit limit is appropriately reduced and the installment rate is increased; for high-risk customers, the credit limit is significantly reduced, the rate is increased, and transaction restrictions are added.

[0108] In order to verify the effectiveness of the differentiated credit strategy, the present invention uses the historical data of the bank in the past two years to conduct a simulation test. The results show that compared with the original unified credit policy, the differentiated credit strategy reduces the non-performing loan rate by about 20% while maintaining the overall return level. The comparison dimensions are shown in Table 1.

[0109] Table 1 Comparison dimensions between the present invention and the prior art

[0110] Comparison Dimensions The present invention Prior art Model building Combining marginal revenue forecast and credit risk forecast Consider only credit risk factors Strategy Development Differentiated credit strategies Unified credit strategy Optimization goals Balancing benefits and risks Single focus on risk control Implementation Effect The overall profit level remains unchanged, and the bad debt rate drops by 20% NPL ratio drops, but earnings suffer

[0111] In general, the method of the present invention realizes differentiated management of credit product customers through the organic combination of marginal profit and credit risk, effectively reduces risk exposure while ensuring overall profit, and shows good application prospects.

[0112] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A C-card strategy method based on marginal benefit algorithm, characterized by: include, Obtaining original customer data and preprocessing it to generate a standardized data set, wherein the original customer data includes customer attribute data, transaction flow data, credit assessment data, and historical default data; Based on the standardized data set, constructing a marginal revenue prediction model and a credit risk prediction model respectively; Using the marginal revenue prediction model and the credit risk prediction model, predicting the marginal revenue and default probability of credit product customers; Formulate differentiated credit granting strategies for credit product customers based on the marginal benefits and the default probability; Verify the effectiveness of the differentiated credit strategy through strategy simulation or historical data, and make optimization adjustments based on the verification results.

2. The C-card strategy method based on marginal benefit algorithm as claimed in claim 1, characterized in that: The construction process of the marginal revenue prediction model is as follows: Using the customer attribute data and transaction flow data in the standardized data set, a two-layer feature selection mechanism is constructed to generate the first feature subset; Calculate the marginal contribution value of each customer's credit product based on the transaction flow data in the standardized data set to generate customer marginal revenue data; Constructing an XGBoost prediction base model, and using the first feature subset and the customer marginal revenue data for training; Using the credit assessment data in the standardized data set, we build a yield correction sub-model and calculate the yield adjustment factor; The XGBoost prediction base model and the revenue correction sub-model are integrated to obtain a marginal revenue prediction model.

3. The C-card strategy method based on marginal benefit algorithm as claimed in claim 2, characterized in that: The construction process of the XGBoost prediction base model is as follows: constructing a first training feature matrix based on the first feature subset, and using the customer marginal revenue data as a first training label; Dividing the first training feature matrix and its corresponding first training label into a first training subset and a first validation subset according to a preset ratio; Constructing a first decision tree, and using the first training subset to calculate the splitting gain of the first decision tree, and determining an optimal splitting feature and a splitting threshold; Iteratively construct a decision tree sequence based on the residual of the previous decision tree until the preset number of iterations is reached; The decision tree sequences are combined to form an XGBoost prediction base model, and the initial marginal revenue prediction value is output.

4. The C-card strategy method based on marginal benefit algorithm as claimed in claim 1, characterized in that: The construction process of the credit risk prediction model is as follows: Using the transaction flow data in the standardized data set to construct a customer transaction association network; Using a community discovery algorithm to divide the customer transaction association network into groups and generate group identifiers; Extracting customer node association features according to the customer transaction association network, and generating group features in combination with the group identifier; Calculating risk transmission characteristics based on the customer transaction association network; Determine whether the customer has defaulted on a contract based on the transaction flow data in the standardized data set, and generate a default mark; Based on the customer node association characteristics, the group characteristics and the risk conduction characteristics, construct a second training feature matrix, and divide the second training feature matrix into a second training subset and a second verification subset according to a preset ratio; Build and train a credit risk prediction model based on the gradient boosting framework.

5. The C-card strategy method based on marginal benefit algorithm as claimed in claim 4, characterized in that: The construction and training of the credit risk prediction model based on the gradient boosting framework includes the following steps: Determine the initial default probability prediction value p0 for all samples; Calculate the negative gradient value of each sample in the second training subset based on the log-likelihood loss function; The negative gradient value is used as the target variable, and a CART decision tree is constructed by combining the customer node association characteristics, group characteristics, and risk conduction characteristics to obtain a risk prediction base learner, and output the t-th round prediction adjustment value Δp t ; The product of the output result of the risk prediction base learner and the preset learning rate is used as the default probability adjustment amount of the current round; Adding the default probability adjustment amount to the default probability forecast value of the previous round to obtain an updated default probability forecast value; Using AUC as the model evaluation indicator, the negative gradient calculation and model iterative update process are repeated until the convergence conditions are met to obtain the final credit risk prediction model.

6. The C-card strategy method based on marginal benefit algorithm as claimed in claim 1, characterized in that: The formulation of a differentiated credit granting strategy for credit product customers includes the following steps: According to the marginal benefit forecast value, determining the marginal benefit group to which the customer belongs, wherein the marginal benefit group includes a high marginal benefit group, a medium marginal benefit group and a low marginal benefit group; Formulate basic credit strategies based on the marginal beneficiary groups; Based on the default probability, the basic credit strategy is adjusted to form a differentiated credit strategy; Output differentiated credit strategy report.

7. The C-card strategy method based on marginal benefit algorithm as claimed in claim 5, characterized in that: The convergence condition includes that the number of iterations of the risk prediction base learner reaches a preset number of iterations, or that the AUC indicator change value on the second verification subset is less than a preset convergence threshold in consecutive preset rounds of iterations of the risk prediction base learner.

8. A C-card strategy system based on a marginal benefit algorithm, based on the C-card strategy method based on a marginal benefit algorithm according to any one of claims 1 to 7, characterized in that: Also includes, The data processing module is used to obtain and preprocess the original customer data to generate a standardized data set; A model building module, used to build a marginal revenue prediction model and a credit risk prediction model based on the standardized data set; A model prediction module, used to predict the marginal profit and default probability of credit product customers by using the marginal profit prediction model and the credit risk prediction model; A strategy formulation module, used to formulate differentiated credit granting strategies for credit product customers based on the marginal benefits and the default probability; The verification module is used to verify the effectiveness of the differentiated credit strategy through strategy simulation or historical data, and to make optimization adjustments based on the verification results.

Images