Strategy grading risk control method and system based on integrated model

Abstract

The invention belongs to the field of loan risk control, discloses a strategy grading risk control method and system based on an integrated model, and is suitable for pre-loan and post-loan links. The method comprises the following steps: constructing and training an integrated model, obtaining rating index characteristics of customers, predicting default probabilities of the customers, and integrating prediction results to obtain default probabilities of all the customers; calculating the validity of a preset risk control rule, and dividing the preset risk control rule in combination with a grading standard; optimizing the hard rule by using a genetic algorithm to determine a final hard rule; constructing a rule scoring model based on the final hard rule and the soft rule; and obtaining rating index characteristics of the current customer, performing calculation by using a rule scoring model, comparing a rule model scoring result of the current customer with a risk decision threshold value configured in advance according to customer group differentiation, and finally executing risk control decision. According to the method, the risk control rule redundancy problem is effectively solved, comprehensive and accurate risk evaluation on customers is realized, and the mistaken killing rate is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of loan risk control, and particularly relates to a strategy hierarchical risk control method and system based on an integrated model. BACKGROUND

[0002] The accuracy and stability of traditional credit risk models are insufficient. Current credit risk assessment models based on logistic regression often predict the default probability of customers based on multiple features, and the model complexity is low, which can easily cause problems such as failure to capture nonlinear relationships, insufficient accuracy and stability.

[0003] The rule effect evaluation is not comprehensive and accurate enough. The existing technology generally evaluates the rule effectiveness based on the rule lift (the actual delinquency rate of the rule hit sample / the actual delinquency rate of the big plate), and the customers who are rejected in the pre-loan approval link have no repayment performance, and a large number of customers who are exited in the post-loan link have no delinquency performance, so the rule lift value is difficult to accurately evaluate the actual effect of each rule, which affects the accuracy and effectiveness of risk management.

[0004] The traditional pre-loan and post-loan link rules are prone to redundancy problems, which highlights the customer mis-killing problem. In the field of customer credit risk assessment, the traditional rule strategy often has redundancy problems, which leads to frequent customer mis-killing.

[0005] The flexibility and decision accuracy of the traditional risk control strategy application are insufficient. The traditional risk control strategy application is often a simple mode of rejection when a single rule is hit, and there is a lack of comprehensive risk assessment of customers. SUMMARY

[0006] The embodiments of the present application provide a strategy hierarchical risk control method and system based on an integrated model to solve the problems in the prior art.

[0007] To have a basic understanding of some aspects of the disclosed embodiments, a brief summary is given below.

[0008] According to a first aspect of the embodiments of the present application, a strategy hierarchical risk control method based on an integrated model is provided.

[0009] In one embodiment, the strategy hierarchical risk control method based on an integrated model comprises:

[0010] building and training a pre-loan integrated model to obtain rating indicator features of customers; combining the trained pre-loan integrated model to predict the default probability of the customers, and integrating the prediction results to obtain the default probability of all customers;

[0011] based on the default probability of all customers, calculating the effectiveness of a preset risk control rule, and combining a hierarchical standard to divide the preset risk control rule into soft rules and hard rules;

[0012] The genetic algorithm is used to optimize the hard rules to determine the final hard rules; and a pre-loan rule scoring model is constructed based on the final hard rules and the soft rules.

[0013] The rating indicator features of the current customer are obtained, the pre-loan rule scoring model is used for calculation to obtain the pre-loan rule model scoring result of the current customer, and the pre-loan rule model scoring result of the current customer is compared with the risk decision threshold pre-configured according to the customer group difference. If the pre-loan rule model scoring result of the current customer exceeds the risk decision threshold, the loan demand of the current customer is rejected.

[0014] In one embodiment, a pre-loan integrated model is constructed and trained to obtain the rating indicator features of the customer; the default probability of the customer is predicted in combination with the trained pre-loan integrated model, and the prediction results are integrated to obtain the default probability of all customers, including the following steps:

[0015] A pre-loan integrated model is constructed based on the rating indicator features of the customer and in combination with the Light-GBM algorithm;

[0016] The hyperparameters of the pre-loan integrated model are optimized by using the grid search to obtain the trained pre-loan integrated model;

[0017] The rating indicator features of the customer are obtained, and the pre-loan integrated model scoring result of the customer is obtained based on the trained pre-loan integrated model;

[0018] The pre-loan integrated model scoring result of the customer is divided into several bins according to the preset interval rule, and the bins are mapped in combination with the penalty coefficient to obtain the default probability of the customer;

[0019] The prediction results are integrated to obtain the default probability of the rejected customer and the default probability of the passed customer, respectively.

[0020] In one embodiment, the effectiveness of the preset risk control rule is the ratio of the average default probability of the rejected customer to the average default probability of the passed customer in the preset risk control rule.

[0021] In one embodiment, the genetic algorithm is used to optimize the hard rules to determine the final hard rules; and a pre-loan rule scoring model is constructed based on the final hard rules and the soft rules, including the following steps:

[0022] An adaptability function is constructed by using the pass rate, the default rate and the profit rate, the final hard rules are obtained by multi-objective optimization of the hard rules;

[0023] A pre-loan rule scoring model is constructed based on the final hard rules and the soft rules;

[0024] According to the rating index characteristics of the customer, whether the customer meets the preset risk control rule is determined, when the rating index characteristics of the customer meet the final hard rule, the customer is directly rejected; when the rating index characteristics of the customer meet the soft rule, a pre-loan rule scoring model is used to calculate the pre-loan rule model score of the customer.

[0025] According to a second aspect of the embodiment of the present application, a strategy grading risk control system based on an integrated model is provided.

[0026] In one embodiment, the strategy grading risk control system based on the integrated model comprises:

[0027] a pre-loan risk prediction module configured to build and train a pre-loan integrated model, acquire the rating index characteristics of the customer, combine the trained pre-loan integrated model to predict the default probability of the customer, and integrate the prediction results to obtain the default probability of all the customers;

[0028] a pre-loan preset risk control rule division module configured to calculate the effectiveness of the preset risk control rule based on the default probability of all the customers, and combine the grading standard to divide the preset risk control rule into a soft rule and a hard rule;

[0029] a pre-loan rule scoring model building module configured to use a genetic algorithm to optimize the hard rule to determine a final hard rule, and build a pre-loan rule scoring model based on the final hard rule and the soft rule;

[0030] a pre-loan strategy grading risk control decision module configured to acquire the rating index characteristics of the current customer, use the pre-loan rule scoring model to calculate the score result of the current customer, compare the pre-loan rule model score result of the current customer with a risk decision threshold pre-configured according to the customer group difference, and if the pre-loan rule model score result of the current customer exceeds the risk decision threshold, reject the loan demand of the current customer.

[0031] According to a third aspect of the embodiment of the present application, a strategy grading risk control method based on an integrated model is provided.

[0032] In one embodiment, the strategy grading risk control method based on the integrated model comprises:

[0033] building and training a post-loan integrated model, acquiring the rating index characteristics of the customer, combining the trained post-loan integrated model to predict the default probability of the customer, and integrating the prediction results to obtain the default probability of all the customers;

[0034] calculating the effectiveness of the preset risk control rule based on the default probability of all the customers, and combining the grading standard to divide the preset risk control rule into a soft rule and a hard rule;

[0035] The genetic algorithm is used to optimize the hard rules, and the final hard rules are determined; and a post-loan rule scoring model is constructed based on the final hard rules and soft rules.

[0036] The rating indicator characteristics of the current customer are obtained, the post-loan rule scoring model is calculated, the post-loan rule model score of the current customer is obtained, and the post-loan rule model score of the current customer is compared with the risk decision threshold configured in advance according to the customer difference, and the post-loan risk control decision is determined and executed according to the comparison result.

[0037] In one embodiment, a post-loan integrated model is constructed and trained, and the rating indicator characteristics of the customer are obtained; the default probability of the customer is predicted in combination with the trained post-loan integrated model, and the prediction results are integrated to obtain the default probability of all customers, including the following steps:

[0038] Based on the rating indicator characteristics of the customer, a post-loan integrated model is constructed in combination with the Light-GBM algorithm;

[0039] The hyperparameters of the post-loan integrated model are optimized by using grid search to obtain the trained post-loan integrated model;

[0040] The rating indicator characteristics of the customer are obtained, and the post-loan integrated model score of the customer is obtained based on the trained post-loan integrated model;

[0041] According to the preset interval rule, the post-loan integrated model score of the customer is divided into several bins, and the bins are mapped in combination with the penalty coefficient to obtain the default probability of the customer;

[0042] The prediction results are integrated to obtain the default probability of the rejected customer and the default probability of the passed customer, respectively.

[0043] In one embodiment, the genetic algorithm is used to optimize the hard rules to determine the final hard rules, including the following steps:

[0044] The pass rate, default rate and profit rate are used to construct a fitness function;

[0045] Based on the fitness function, the hard rules are calculated to obtain the fitness value of the hard rules;

[0046] The affinity crossover function is used to perform affinity crossover and mutation on the hard rules to obtain the offspring of the hard rules;

[0047] The offspring of the hard rules are elitist reserved and selected according to the fitness value of the hard rules, the next generation subset is generated in combination with roulette, and the final hard rules are determined by iterative optimization.

[0048] In one embodiment, the post-loan rule model score result of the current client is compared with the risk decision threshold value configured in advance according to the customer group difference, and a post-loan risk control decision is determined and executed according to the comparison result, including:

[0049] The risk decision threshold value is configured according to the customer group difference, and is divided into a first risk decision threshold value and a second risk decision threshold value, wherein the first risk decision threshold value is greater than the second risk decision threshold value;

[0050] If the post-loan rule model score result of the current client is greater than or equal to the first risk decision threshold value, a first post-loan preset risk control decision is determined and executed;

[0051] If the post-loan rule model score result of the current client is greater than or equal to the second risk decision threshold value and less than the first risk decision threshold value, a second post-loan preset risk control decision is determined and executed.

[0052] According to a fourth aspect of the embodiment of the present application, a strategy grading risk control system based on an integrated model is provided.

[0053] In one embodiment, the strategy grading risk control system based on the integrated model includes:

[0054] A post-loan risk prediction module is configured to build and train a post-loan integrated model, obtain rating index features of a client, combine the trained post-loan integrated model to predict a default probability of the client, and integrate the prediction result to obtain a default probability of all clients;

[0055] A post-loan preset risk control rule division module is configured to calculate the effectiveness of a preset risk control rule based on the default probability of all clients, and divide the preset risk control rule into a soft rule and a hard rule according to a grading standard;

[0056] A post-loan rule scoring model construction module is configured to optimize the hard rule by using a genetic algorithm to determine a final hard rule, and construct a post-loan rule scoring model based on the final hard rule and the soft rule;

[0057] A post-loan strategy grading risk control decision module is configured to obtain rating index features of a current client, calculate a post-loan rule model score result of the current client by using the post-loan rule scoring model, compare the post-loan rule model score result of the current client with a risk decision threshold value configured in advance according to the customer group difference, and determine and execute a post-loan risk control decision according to the comparison result.

[0058] According to a fifth aspect of the embodiment of the present application, a computer device is provided.

[0059] In some embodiments, the computer device includes a memory and a processor, the memory stores a computer program, and the processor implements the steps of the above method when executing the computer program.

[0060] According to a sixth aspect of the embodiments of the present application, a computer readable storage medium is provided.

[0061] In one embodiment, the computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the steps of the above method.

[0062] The technical solutions provided by the embodiments of the present application can have the following beneficial effects:

[0063] 1. The present application can improve the accuracy and stability of model prediction by simplifying modeling features, reducing noise, and automatically optimizing hyperparameters. In the case of rejecting customers, the default probability of the customer can be accurately calculated to provide a scientific basis for decision-making. At the same time, by measuring the average PD of new customer groups, the risk level of new customer groups can be accurately predicted in advance, providing a basis for risk control decisions.

[0064] 2. The rule effectiveness index designed in the present application compares the average default probability of rejected customers and passed customers, which makes up for the blind spots of traditional lift indexes, accurately measures the actual effect of each rule, and ensures the effectiveness of the risk control strategy.

[0065] 3. The present application identifies and removes redundant rules through the rule effectiveness index and optimizes the rule combination using a genetic algorithm, which reduces the risk of mistakenly rejecting customers, significantly reduces the number of rules, significantly reduces the default probability of passed customers, significantly improves the pass rate, and greatly improves the efficiency of the risk control system and customer experience.

[0066] 4. The present application changes the traditional simple and rough mode of customer rejection based on a single rule, realizes risk refinement grading according to the customer rule scoring model score, and thus realizes all-round assessment of customer credit risk, minimizes customer misidentification, minimizes the default probability of approved customers, and maximizes the pass rate.

[0067] 5. The rejection threshold and rule weight in the present application can be flexibly adjusted, supporting differentiated configuration of rejection thresholds for customer groups, operating institutions, customer managers, and industries, which is conducive to optimizing resource allocation, risk management flexibility, and refinement. By regularly monitoring the rule effectiveness of each rule, the rule weight can be flexibly adjusted, and the rule weight has simple and intuitive interpretability.

[0068] 6. The risk assessment of loan customers is a dynamic process, and the credit status and behavior of customers will change over time. Through the dynamic adjustment of rules and the change of customer risk level, the present invention can flexibly adjust the risk control measures and effectively respond to the fluctuations in customer risk. The present invention can solve the problem of evaluating the effectiveness of loan warning rules, realize real-time monitoring of rule effectiveness and rule iteration, improve the accuracy of loan warning exit risk signals, improve the efficiency of loan warning exit risk control systems, reduce unnecessary loan customer misidentification, improve the overall experience of customers, and reduce customer churn rate.

[0069] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the present invention. BRIEF DESCRIPTION OF DRAWINGS

[0070] The accompanying drawings incorporated in and forming a part of the specification illustrate embodiments consistent with the present invention and, together with the description, serve to explain the principles of the invention.

[0071] Figure 1 is a pre-loan flowchart of the integrated model-based strategy grading risk control method according to an exemplary embodiment;

[0072] Figure 2 is a pre-loan principle block diagram of the integrated model-based strategy grading risk control system according to an exemplary embodiment;

[0073] Figure 3 is a post-loan flowchart of the integrated model-based strategy grading risk control method according to an exemplary embodiment;

[0074] Figure 4 is a post-loan principle block diagram of the integrated model-based strategy grading risk control system according to an exemplary embodiment;

[0075] Figure 5 is a structural schematic diagram of a computer device according to an exemplary embodiment;

[0076] Figure 6 is a general flowchart of the integrated model-based strategy grading risk control method according to an exemplary embodiment. DETAILED DESCRIPTION

[0077] The following description and drawings are illustrative of specific embodiments thereof and are not intended to limit the scope of the embodiments. Parts and features of some embodiments can be included or substituted in or for parts and features of other embodiments. The scope of the embodiments encompassed herein includes the whole scope of the claims together with all available equivalents of the claims. In this document, the terms "first", "second", etc. are used merely to distinguish one element from another, and do not require or imply any actual relationship or order between the elements. In fact, the first element can be referred to as the second element, and vice versa. Also, the terms "comprises", "comprising", or any other variations thereof are intended to cover a non-exclusive inclusion, such that a structure, device, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such structure, device, or apparatus. Without further limitation, an element defined by an "includes a" statement does not exclude the presence of additional identical elements in the structure, device, or apparatus that includes the element. Various embodiments are described in progressive stages, each of which focuses on the differences from other embodiments, and the same or similar parts between various embodiments can be referred to each other.

[0078] In this document, the terms "longitudinal", "lateral", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", and the like indicate the orientation or positional relationship shown in the drawings, and are used only for the convenience of description and simplification of the description herein, and do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application. In the description herein, unless otherwise specified and limited, the terms "mount", "connect", "connection" should be interpreted broadly, for example, it can be a mechanical connection or an electrical connection, it can be a communication between two elements inside, it can be a direct connection or an indirect connection through an intermediate medium, and the specific meaning of the above terms can be understood by the person skilled in the art according to the specific circumstances.

[0079] In this document, the term "multiple" means two or more, unless otherwise specified.

[0080] In this document, the character " / " represents an "or" relationship between the objects before and after it. For example, A / B means A or B.

[0081] In this document, the term "and / or" is a description of the relationship between the objects, which means that there can be three relationships. For example, A and / or B means that there are three relationships of A or B, or A and B.

[0082] It should be understood that although the steps in the flowchart are shown in a sequential order following the arrows, the steps are not necessarily executed in the order shown by the arrows. Unless otherwise explicitly stated herein, there is no strict order requirement for the execution of the steps, and the steps can be executed in other orders. Moreover, at least some of the steps in the figure can include multiple sub-steps or multiple stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution of the sub-steps or stages can not necessarily be sequential, but can be performed alternately or alternately with at least part of other steps or sub-steps or stages of other steps.

[0083] Each module in the device or system of the present application can be implemented wholly or partially by software, hardware, and a combination thereof. The above modules can be embedded in or independent of the processor in the computer device in hardware form, or stored in the memory in the computer device in software form, so as to be called and executed by the processor to perform the operations corresponding to the above modules.

[0084] In the case of no conflict, the embodiments in the present application and the features in the embodiments can be combined with each other.

[0085] Figure 1 And Figure 6 An embodiment of the integrated model-based policy grading risk control method of the present application is shown.

[0086] In this optional embodiment, the integrated model-based policy grading risk control method comprises:

[0087] S101, constructing and training a pre-loan integrated model to obtain rating index features of a customer; combining the trained pre-loan integrated model to predict the default probability of the customer, and integrating the prediction results to obtain the default probability of all customers;

[0088] S102, based on the default probability of all customers, calculating the effectiveness of a preset risk control rule, and combining the grading standard to divide the preset risk control rule into a soft rule and a hard rule;

[0089] S103, using a genetic algorithm to optimize the hard rule to determine the final hard rule; based on the final hard rule and the soft rule, constructing a pre-loan rule scoring model;

[0090] S104, obtaining the rating index features of the current customer, using the pre-loan rule scoring model to calculate the pre-loan rule model score of the current customer, comparing the pre-loan rule model score of the current customer with the risk decision threshold pre-configured according to the customer group difference, and if the pre-loan rule model score of the current customer exceeds the risk decision threshold, rejecting the loan demand of the current customer.

[0091] In this optional embodiment, a pre-loan integrated model is constructed and trained to obtain the rating indicator features of the customer; the pre-loan integrated model is combined with the trained pre-loan integrated model to predict the default probability of the customer, and the prediction results are integrated to obtain the default probability of all customers, including the following steps:

[0092] Based on the rating indicator features of the customer, a pre-loan integrated model is constructed in combination with the Light-GBM algorithm;

[0093] The hyperparameters of the pre-loan integrated model are optimized using grid search to obtain the trained pre-loan integrated model;

[0094] The rating indicator features of the customer are obtained, and the pre-loan integrated model score of the customer is obtained based on the trained pre-loan integrated model;

[0095] According to the preset interval rule, the pre-loan integrated model score of the customer is divided into several bins, and the bins are mapped in combination with the penalty coefficient to obtain the default probability of the customer;

[0096] The prediction results are integrated to obtain the default probability of the rejected customers and the default probability of the passed customers, respectively.

[0097] In this optional embodiment, the effectiveness of the preset risk control rule is the ratio of the average default probability of the rejected customers and the average default probability of the passed customers in the preset risk control rule.

[0098] In this optional embodiment, the genetic algorithm is used to optimize the hard rules to determine the final hard rules; based on the final hard rules and the soft rules, a pre-loan rule scoring model is constructed, including the following steps:

[0099] A fitness function is constructed using the pass rate, default rate, and profit rate, and the final hard rules are obtained by multi-objective optimization of the hard rules;

[0100] Based on the final hard rules and the soft rules, a pre-loan rule scoring model is constructed;

[0101] According to the rating indicator features of the customer, the hit situation of the customer in the preset risk control rule is determined; when the rating indicator features of the customer hit the final hard rules, the customer is directly rejected; when the rating indicator features of the customer hit the soft rules, the pre-loan rule model score of the customer is calculated using the pre-loan rule scoring model.

[0102] It needs to be explained that in the pre-loan approval link, the rules are classified according to the effectiveness of the preset risk control rule (hard rules and soft rules), and the pre-loan rule application method is differentiated. According to the pre-loan rule model score, the threshold is configured for the customer group, and the risk control method is differentiated. The pre-loan strategy classification risk control method of the integrated model includes the following steps:

[0103] I. Pre-loan integrated model modeling and PD (i.e., default probability) calculation.

[0104] As shown in Table 1, key scoring dimensions (i.e., customer rating indicator features) are selected as inputs of the pre-loan integrated model, which are derived from multi-dimensional scoring of pre-loan customers, including:

[0105] Table 1: Customer rating indicator features

[0106]

[0107] Using the LightGBM integrated learning method, the prediction results of multiple scores are integrated, which not only makes full use of a small amount of high-value score information, but also improves the prediction performance and stability, avoiding overfitting problems caused by too many features.

[0108] Using the trained integrated model, the default probability of each customer is predicted; at the same time, the customer score output by the model is divided into N bins according to the pre-set rules, and is mapped to the actual default probability, ensuring that the predicted PD is highly matched with the actual default rate, thereby providing an accurate basis for subsequent rule application. For rejected customers, a penalty coefficient (e.g., 1.2) can be set, and the default probability of all rejected customers is magnified (default probability * penalty coefficient of rejected customers) to prevent underestimating the default probability of rejected customers. As shown in Table 2, the mapping results of pre-loan binning are shown.

[0109] Table 2: Pre-loan binning mapping results

[0110] Integrated model score binning Actual default probability (0，10] 20％ (10，20] 10％ ... ... (90，100] 0.5％

[0111] II. Pre-loan rule effectiveness evaluation.

[0112] The effectiveness of the pre-set risk control rule is calculated, which is defined as the ratio of the average default probability of customers in the pre-set risk control rule to the average default probability of passed customers. The higher the effectiveness ratio, the better the rule distinguishes between high-risk and low-risk customers.

[0113] III. Rule effectiveness classification.

[0114] As shown in Table 3, according to the rule effectiveness, the pre-set risk control rules are classified into two categories:

[0115] Table 3: Classification of pre-set risk control rules

[0116]

[0117] In Table 3, N1 represents the first pre-set risk control rule threshold (applicable to the pre-loan link); N2 represents the second pre-set risk control rule threshold (applicable to the pre-loan link).

[0118] The application is based on the genetic algorithm optimization rule combination of multi-objective decision, introduces different objectives (pass rate, default rate, profit rate) into the fitness function, finds the optimal combination of rules through multi-objective optimization, adds an adaptive mechanism to dynamically adjust algorithm parameters, and maximizes the institutional revenue.

[0119] The genetic algorithm is used to optimize the risk control rule set, different objectives (pass rate, default rate, profit rate) are introduced into the fitness function, and the optimal combination of rules (i.e., preset risk control rules) is found through multi-objective optimization. An adaptive mechanism is added to dynamically adjust algorithm parameters, such as automatically increasing or decreasing mutation rate and crossover rate according to the current population diversity, to speed up the convergence speed and avoid falling into local optimum. It is applicable to pre-loan and post-loan links, and the specific steps are as follows:

[0120] 1. Fitness function setting.

[0121] The pass rate, default rate, and profit rate of each rule combination are calculated, the fitness function is set, and the pass rate and profit rate are maximized and the default probability PD is minimized. The fitness function can configure a rule number reward and punishment mechanism, configure the expected number of rules (e.g., 100), and perform deviation punishment. The expression formula of the fitness function is:

[0122] Fitness(g) = w1*pass rate-w2*default rate PD+w3*profit rate;

[0123] In the formula, w1, w2, and w3 represent the weight parameters of the pass rate, default rate, and profit rate, respectively; and Fitness(g) represents the fitness function. The fitness function can configure demand restriction conditions, such as the default rate must be lower than 2%, and the fitness is 0 if the restriction condition is not met.

[0124] The single customer profit rate can be approximately predicted as: loan interest rate-default probability PD-fund cost rate-operating cost rate, and the profit rate of the full amount of customers is amount weighted according to the expected usage amount (amount*usage rate).

[0125] 2. Affinity crossover and mutation.

[0126] The affinity crossover function generates offspring according to the union of the parents and ensures that the fixed rules (rules that cannot be deleted) are always retained; the mutation function randomly adds or deletes non-fixed rules.

[0127] 3. Elite reservation and selection.

[0128] The highest fitness combination of each generation is reserved, and roulette selection is used to generate offspring.

[0129] 4. Output.

[0130] The final output is the best rule combination (rule name set) and the best fitness.

[0131] Four, pre-loan rule scoring model.

[0132] Based on the final hard rules and soft rules, a pre-loan rule scoring model is constructed, and the calculation formula of the pre-loan rule scoring model is:

[0133]

[0134] In the formula, W i represents the weight of the i-th rule; X i represents the hit of the customer in the rule; and n represents the number of rules.

[0135] As shown in Table 4, an example of the effect of the pre-loan rule scoring model is shown.

[0136] Table 4: Example of pre-loan rule scoring model effect

[0137] Rule score model score interval Average PD Actual 30 day+ delinquency rate 0-1 1.36％ 1.30％ 1-2 1.84％ 1.90％ 2-3 3.09％ 3.00％ 3-4 3.87％ 3.86％ 4-5 4.95％ 4.91％ 5-6 5.00％ 4.90％ 6-7 5.17％ 6.53％ 7-8 6.00％ 7.01％ 8-9 7.00％ 7.34％ 9-11 7.20％ 7.31％ 11-13 8.00％ 8.86％ 13-16 8.24％ 9.14％ 16-18 9.90％ 9.76％ 18-20 11.20％ 12.12％ 20-25 14.00％ 13.80％ 25-30 15.00％ 15.00％ 30-35 16.00％ 16.00％ 35-40 17.20％ 17.00％ 40-60 18.10％ 18.00％ 60-80 18.70％ 19.00％ 80-inf 21.00％ 20.00％

[0138] Five, pre-loan risk decision.

[0139] According to the comparison between the pre-loan rule model scoring result and the preset rejection threshold (i.e. risk decision threshold), if the score exceeds the rejection threshold, the customer will be rejected, thereby realizing risk grading and decision automation. As shown in Table 5, the rejection threshold can be configured according to the difference of the customer group. For example, the rejection threshold of the preferred customer can be relaxed to improve the pass rate and improve the flexibility of risk control decision. The access threshold can be differentiated for branch banks, branch banks, customer managers, industries and other sub-groups to optimize resource allocation efficiency and asset quality.

[0140] Table 5: Example of different threshold setting effect

[0141] Customer stratification Rule score model score rejection threshold Through customer average PD Pass rate Pass rate lift Super prime customer 5 1％ 56％ 23％ Prime customer 3 1.5％ 20％ 5％ Near prime customer 2 2％ 5％ 2％ Subprime customer 1 2.5％ 2％ 1％

[0142] Figure 2 An embodiment of the strategy grading risk control system based on the integrated model of the application is shown.

[0143] In this optional embodiment, the strategy grading risk control system based on the integrated model comprises:

[0144] The pre-loan risk prediction module 201 is used to construct and train the pre-loan integrated model to obtain the rating index features of the customer; the pre-loan integrated model after training is combined to predict the default probability of the customer, and the prediction results are integrated to obtain the default probability of all customers.

[0145] The pre-loan preset risk control rule division module 202 is configured to calculate the effectiveness of the preset risk control rule based on the default probability of all customers, and divide the preset risk control rule into soft rules and hard rules in combination with the grading standard;

[0146] The pre-loan rule scoring model construction module 203 is configured to optimize the hard rules by using a genetic algorithm, determine the final hard rules, and construct a pre-loan rule scoring model based on the final hard rules and the soft rules.

[0147] The pre-loan strategy grading risk control decision module 204 is configured to obtain the rating index features of a current customer, calculate the pre-loan rule model score of the current customer by using the pre-loan rule scoring model, and compare the pre-loan rule model score of the current customer with a risk decision threshold pre-configured according to customer group differentiation, and if the pre-loan rule model score of the current customer exceeds the risk decision threshold, the loan demand of the current customer is rejected.

[0148] Figure 3 and Figure 6 An embodiment of the strategy grading risk control method based on an integrated model is shown.

[0149] In this optional embodiment, the strategy grading risk control method based on an integrated model comprises:

[0150] S301, constructing and training a post-loan integrated model to obtain the rating index features of a customer, combining the trained post-loan integrated model to predict the default probability of the customer, and integrating the prediction results to obtain the default probability of all customers;

[0151] S302, calculating the effectiveness of the preset risk control rule based on the default probability of all customers, and dividing the preset risk control rule into soft rules and hard rules in combination with the grading standard;

[0152] S303, optimizing the hard rules by using a genetic algorithm, determining the final hard rules, and constructing a post-loan rule scoring model based on the final hard rules and the soft rules;

[0153] S304, obtaining the rating index features of a current customer, calculating the post-loan rule model score of the current customer by using the post-loan rule scoring model, comparing the post-loan rule model score of the current customer with a risk decision threshold pre-configured according to customer group differentiation, and determining and executing a post-loan risk control decision according to the comparison result.

[0154] In this optional embodiment, constructing and training a post-loan integrated model to obtain the rating index features of a customer, combining the trained post-loan integrated model to predict the default probability of the customer, and integrating the prediction results to obtain the default probability of all customers comprise the following steps:

[0155] Based on the rating index features of the customer, and combined with the Light-GBM algorithm, a post-loan integrated model is constructed;

[0156] The hyperparameters of the post-loan integrated model are optimized by using grid search, and the trained post-loan integrated model is obtained;

[0157] The rating index features of the customer are obtained, and based on the trained post-loan integrated model, the post-loan integrated model score result of the customer is obtained;

[0158] According to the preset interval rule, the post-loan integrated model score result of the customer is divided into several bins, and the bins are mapped combined with the penalty coefficient to obtain the default probability of the customer;

[0159] The prediction results are integrated to obtain the default probability of the rejected customers and the default probability of the passed customers, respectively.

[0160] In this optional embodiment, the genetic algorithm is used to optimize the hard rules to determine the final hard rules, including the following steps:

[0161] The pass rate, default rate, and profit rate are used to construct the fitness function;

[0162] Based on the fitness function, the hard rules are calculated to obtain the fitness value of the hard rules;

[0163] The affinity crossover function is used to perform affinity crossover and mutation on the hard rules to obtain the offspring of the hard rules;

[0164] According to the fitness value of the hard rules, the offspring of the hard rules are elitist reserved and selected, combined with roulette to generate the next generation subset, and the final hard rules are determined through iterative optimization.

[0165] In this optional embodiment, the post-loan rule model score result of the current customer is compared with the risk decision threshold value pre-configured according to the customer group differentiation, and the post-loan risk control decision is determined and executed according to the comparison result, including:

[0166] The risk decision threshold value is configured according to the customer group differentiation, and is divided into a first risk decision threshold value and a second risk decision threshold value, wherein the first risk decision threshold value is greater than the second risk decision threshold value;

[0167] If the post-loan rule model score result of the current customer is greater than or equal to the first risk decision threshold value, the first post-loan preset risk control decision is determined and executed;

[0168] If the post-loan rule model score result of the current customer is greater than or equal to the second risk decision threshold value and less than the first risk decision threshold value, the second post-loan preset risk control decision is determined and executed.

[0169] It needs to be explained that the traditional early warning strategy rule in the post-loan stage may also have redundancy and inefficiency problems. Moreover, for a large number of customers who have no overdue performance after exiting the stop amount, there is a lack of effective indicators to evaluate the effectiveness of the rules. Therefore, it is impossible to judge the actual effect of the post-loan early warning rule, which may lead to a large number of customer miskills. In the post-loan link, the rules are classified according to the rule effectiveness (hard rules and soft rules), and the rule application method is differentiated. According to the rule model score, the threshold is configured for the risk control method of customer differentiation. The post-loan strategy grading risk control method of the integrated model includes the following steps:

[0170] I. Post-loan integrated model modeling and PD (i.e., default probability) calculation.

[0171] The LightGBM integrated learning method can be used to integrate multiple post-loan early warning model scores, customer value contribution model scores, customer hierarchical levels, multi-head levels, and other post-loan customer evaluation results, improve the prediction performance and stability, and avoid the overfitting problem caused by too many features. The trained post-loan integrated model is used to predict the default probability (PD) of each customer; at the same time, the customer score output by the model is divided into N bins according to the preset rules, and is mapped to the actual default probability, ensuring that the predicted PD is highly matched with the actual default rate, thereby providing an accurate basis for subsequent rule application. If there are no multiple post-loan early warning model scores, a machine learning early warning model based on post-loan features can be used instead of the post-loan integrated model.

[0172] II. Post-loan rule effectiveness evaluation.

[0173] The effectiveness of the post-loan preset risk control rule is calculated, which is defined as the ratio of the average default probability of the customers in the preset risk control rule to the average default probability of the customers passing the rule. Invalid rules with a rule effectiveness ratio less than 1 should be removed in a timely manner.

[0174] III. Rule effectiveness classification.

[0175] As shown in Table 6, according to the rule effectiveness, the rules are classified into different levels (such as hard rules and soft rules) according to the rule effectiveness:

[0176] Table 6: Classification of preset risk control rules

[0177]

[0178] In Table 6, K1 represents the first preset risk control rule threshold (applicable to the post-loan link); K2 represents the second preset risk control rule threshold (applicable to the post-loan link).

[0179] Using genetic algorithm to optimize the combination of risk control rules, first, set the fitness function to calculate the pass rate and default rate of each rule combination, set the fitness function to maximize the pass rate and minimize the default probability PD. The fitness function can configure the rule number reward and punishment mechanism, configure the expected number of rules (for example 100), and perform bias penalty. Then, affinity crossover and mutation are performed to generate offspring according to the union of parents. Next, elite reservation and selection are performed, and the highest fitness combination is reserved in each generation, and roulette is used to select to produce offspring. Finally, the best rule combination and the best fitness are output.

[0180] Four, post-loan rule scoring model.

[0181] Based on the final hard and soft rules, a post-loan rule scoring model is constructed, and the calculation formula of the post-loan rule scoring model is:

[0182]

[0183] In the formula, W i represents the weight of the ith rule; X i represents the hit of the customer in the rule; and n represents the number of rules.

[0184] Five, post-loan risk decision.

[0185] According to the comparison between the customer post-loan rule model score and the preset threshold, risk grading and decision automation are realized. The preset threshold can be configured differently for different groups of customers to improve the flexibility of risk control decisions.

[0186] As shown in Table 7, it is an example of post-loan risk decision. According to the comparison between the post-loan rule scoring model score and the risk decision threshold, risk grading and decision automation are realized.

[0187] Table 7: Post-loan risk grading decision example

[0188]

[0189] In Table 7, S1 represents the first risk decision threshold; and S2 represents the second risk decision threshold.

[0190] The risk assessment of post-loan customers is a dynamic process, and the credit status and behavior of customers will change over time. Through dynamic adjustment of rules and changes in customer risk levels, risk control measures can be flexibly adjusted to effectively respond to fluctuations in customer risk.

[0191] The combination of post-loan customer stratification and rule grading allows for the use of lightweight risk control strategies for low-risk customers, thereby reducing unnecessary customer false positives, improving the overall customer experience, and reducing customer churn.

[0192] Figure 4An embodiment of the integrated model-based policy hierarchical risk control system of the application is shown.

[0193] In this optional embodiment, the integrated model-based policy hierarchical risk control system comprises:

[0194] The post-loan risk prediction module 401 is configured to build and train a post-loan integrated model, obtain rating index features of a customer, combine the trained post-loan integrated model to predict the default probability of the customer, and integrate the prediction results to obtain the default probability of all customers.

[0195] The post-loan preset risk control rule division module 402 is configured to calculate the effectiveness of the preset risk control rule based on the default probability of all customers, and divide the preset risk control rule into soft rules and hard rules in combination with the hierarchical standard.

[0196] The post-loan rule scoring model construction module 403 is configured to optimize the hard rules by using a genetic algorithm to determine the final hard rules, and construct a post-loan rule scoring model based on the final hard rules and the soft rules.

[0197] The post-loan policy hierarchical risk control decision module 404 is configured to obtain the rating index features of a current customer, calculate the post-loan rule scoring model score of the current customer by using the post-loan rule scoring model, compare the post-loan rule scoring model score of the current customer with a risk decision threshold pre-configured according to customer group differentiation, and determine and execute a post-loan risk control decision according to the comparison result.

[0198] It should be explained that the present application remolds the existing traditional pre-loan and post-loan risk control system, comprehensively and accurately evaluates the rule effectiveness, effectively solves the rule redundancy problem, significantly reduces the customer default rate, and significantly improves the pass rate. The present application realizes comprehensive and accurate risk evaluation of customers and reduces the false kill rate.

[0199] 1. Improve model accuracy: by simplifying modeling features, reducing noise, and through automatic hyperparameter tuning, improve the accuracy and stability of model prediction.

[0200] 2. Accurately calculate default probability: among the rejected customers, the default probability can be accurately calculated to provide a scientific basis for decision-making; at the same time, by measuring the average PD of new customer groups, the risk level of new customer groups can be accurately predicted in advance to provide a basis for risk control decision-making.

[0201] 3. Evaluate the rule effectiveness accurately: the existing technology generally assesses the rule effectiveness based on the rule lift (the overdue rate of the rule hit sample / the average overdue rate of the market), which uses the risk performance of the passed customers. The customers rejected in the pre-loan approval link have no repayment performance, and the large number of customers exited in the post-loan link have no overdue performance, so the rule lift value is difficult to accurately evaluate the actual effect of each rule, which affects the accuracy and effectiveness of risk management.

[0202] The rule effectiveness index designed in the application compares the average default probability of the rejected customers and the passed customers, makes up for the blind spot of the traditional lift index, accurately measures the actual effect of each rule, and ensures the effectiveness of the risk control strategy.

[0203] 4. Solve the rule redundancy problem and reduce the customer rejection: the rule redundancy is identified and removed through the rule effectiveness index, and the genetic algorithm is used to optimize the rule combination, which reduces the risk of mistakenly rejecting customers, can greatly reduce the number of rules, greatly reduces the default probability PD of the passed customers, greatly improves the pass rate, and greatly improves the efficiency of the risk control system and the customer experience. The genetic algorithm optimizes the rule combination based on multi-objective decision, introduces different targets (pass rate, default rate, profit rate) into the fitness function, and finds the optimal combination of rules through multi-objective optimization. The adaptive mechanism is added to dynamically adjust the algorithm parameters, and the maximum institutional revenue is obtained.

[0204] As shown in Table 8, it is an effect example of the pre-loan grading risk control decision.

[0205] Table 8: Effect example of pre-loan grading risk control decision

[0206]

[0207] 5. Rule scoring model realizes all-round accurate risk assessment: bad customers often hit many rules at the same time, and customers hitting only a few rules have a high probability of being mistakenly rejected. The application changes the traditional simple and rough mode of customer rejection according to a single rule, realizes the risk fine grading according to the customer rule model score, and thus realizes the all-round assessment of the customer credit risk, minimizes the customer rejection, realizes the minimization of the default probability of the approved customers, and maximizes the pass rate and profit rate.

[0208] 6. The rule scoring model is flexible in application and has intuitive interpretability; the rejection threshold and rule weight can be flexibly adjusted, supports the differentiated configuration of the rejection threshold of the customer group, operating institution, customer manager and industry, and is beneficial to optimizing the resource allocation, risk management flexibility and refinement. The rule effectiveness of each rule is monitored regularly, and the rule weight is flexibly adjusted. It has simple and intuitive interpretability: the contribution degree of a single rule of the rule scoring model is clear.

[0209] 7. Dynamic assessment of post-loan risks, rule iteration mechanism and precise risk signals: The risk assessment of post-loan customers is a dynamic process, and the credit status and behavior of customers will change over time. Through the dynamic adjustment of rules and the change of customer risk levels, risk control measures can be flexibly adjusted to effectively respond to fluctuations in customer risks. It can solve the problem of evaluating the effectiveness of post-loan early warning rules, realize real-time monitoring of rule effectiveness and rule iteration. Improve the precision of post-loan early warning exit risk signals, significantly reduce the number of post-loan early warning and exit work orders, and significantly improve the efficiency of post-loan early warning and exit risk control systems. Reduce unnecessary post-loan customer kills, improve the overall customer experience, and reduce customer churn.

[0210] In an embodiment, a computer device, which can be a server, is provided, and an internal structure diagram of the computer device can be as shown in Figure 5 The computer device includes a processor, a memory and a network interface connected through a system bus. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for running the operating system and the computer program in the non-volatile storage medium. The database of the computer device is configured to store static information and dynamic information data. The network interface of the computer device is configured to communicate with an external terminal through a network connection. The computer program is executed by the processor to implement the steps in the above method embodiments.

[0211] Those skilled in the art can understand that Figure 5 The structure shown in the above

[0212] In addition, the present application also provides a computer device including a memory and a processor, the memory stores a computer program, and the processor implements the steps in the above method embodiments when executing the computer program.

[0213] In addition, the present application also provides a computer readable storage medium having a computer program stored thereon, and the computer program is executed by a processor to implement the steps in the above method embodiments.

[0214] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage medium, and when executed, can include the processes of the above-mentioned embodiments. Any reference to memory, storage, database or other medium used in each embodiment of the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory or optical memory, etc. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0215] The present application is not limited to the structures described above and shown in the drawings, and various modifications and changes can be made without departing from the scope thereof. The scope of the present application is only limited by the appended claims.

Claims

1. A strategy-based hierarchical risk control method based on an ensemble model, characterized in that, The method includes: Build and train a pre-loan integrated model to obtain customer rating indicator features; combine the trained pre-loan integrated model to predict the customer's default probability, and integrate the prediction results to obtain the default probability of all customers. Based on the default probability of all customers, the effectiveness of the preset risk control rules is calculated, and combined with the grading standard, the preset risk control rules are divided into soft rules and hard rules. A genetic algorithm is used to optimize the hard rules and determine the final hard rules; based on the final hard rules and soft rules, a pre-loan rule scoring model is constructed. The system obtains the current customer's rating indicators and uses a pre-loan rule scoring model to calculate the current customer's pre-loan rule model score. The current customer's pre-loan rule model score is then compared with the risk decision threshold pre-configured based on customer group differentiation. If the current customer's pre-loan rule model score exceeds the risk decision threshold, the current customer's loan request is rejected.

2. The strategy-based hierarchical risk control method based on an integrated model according to claim 1, characterized in that, The pre-loan integrated model is constructed and trained to obtain customer rating indicator features; The process of predicting customer default probabilities using a trained pre-loan ensemble model and integrating the prediction results to obtain the default probabilities for all customers includes the following steps: Based on customer rating indicators and combined with the Light-GBM algorithm, a pre-loan integrated model is constructed. The hyperparameters of the pre-loan ensemble model are optimized using grid search to obtain the trained pre-loan ensemble model. Obtain customer rating indicator features and, based on the trained pre-loan integrated model, obtain the customer's pre-loan integrated model score result; According to the preset interval rules, the customer's pre-loan integrated model scoring results are divided into several bins, and the bins are mapped in combination with the penalty coefficient to obtain the customer's default probability. By integrating the prediction results, the default probability of rejecting customers and the default probability of passing customers are obtained separately.

3. The strategy-based hierarchical risk control method based on an integrated model according to claim 2, characterized in that, The effectiveness of the preset risk control rule is the ratio of the average probability of default for rejected customers to the average probability of default for approved customers in the preset risk control rule.

4. The strategy-based hierarchical risk control method based on an integrated model according to claim 1, characterized in that, The process of using a genetic algorithm to optimize the hard rules and determine the final hard rules, and then constructing a pre-loan rule scoring model based on the final hard rules and soft rules, includes the following steps: A fitness function is constructed using pass rate, default rate, and profit margin. The final hard rules are obtained by performing multi-objective optimization on the hard rules. Based on the final hard rules and soft rules, a pre-loan rule scoring model is constructed; The system determines a customer's compliance with preset risk control rules based on their rating indicators. If a customer's rating indicators match the final hard rules, the customer is rejected directly. If a customer's rating indicators match the soft rules, the system calculates the customer's pre-loan rule scoring model score using the pre-loan rule scoring model.

5. A strategy-based hierarchical risk control system based on an integrated model, characterized in that, The system includes: The pre-loan risk prediction module is used to build and train the pre-loan integrated model to obtain customer rating indicator characteristics; combine the trained pre-loan integrated model to predict the customer's default probability, and integrate the prediction results to obtain the default probability of all customers. The pre-loan risk control rule classification module is used to calculate the effectiveness of the pre-set risk control rules based on the default probability of all customers, and to classify the pre-set risk control rules into soft rules and hard rules in combination with the grading standard. The pre-loan rule scoring model construction module is used to optimize the hard rules using a genetic algorithm to determine the final hard rules; based on the final hard rules and soft rules, a pre-loan rule scoring model is constructed. The pre-loan strategy tiered risk control decision module is used to obtain the rating indicator characteristics of the current customer, calculate the pre-loan rule scoring model score of the current customer using the pre-loan rule scoring model, compare the current customer's pre-loan rule scoring model score with the risk decision threshold configured in advance according to the customer group differentiation, and reject the current customer's loan request if the current customer's score exceeds the risk decision threshold.

6. A strategy-based hierarchical risk control method based on an integrated model, characterized in that, The method includes: Build and train a post-loan integrated model to obtain customer rating indicator features; combine the trained post-loan integrated model to predict the customer's default probability, and integrate the prediction results to obtain the default probability of all customers. Based on the default probability of all customers, the effectiveness of the preset risk control rules is calculated, and combined with the grading standard, the preset risk control rules are divided into soft rules and hard rules. A genetic algorithm is used to optimize the hard rules and determine the final hard rules; based on the final hard rules and soft rules, a post-loan rule scoring model is constructed. Obtain the current customer's rating indicator characteristics, use the post-loan rule scoring model to calculate the current customer's post-loan rule model score result, compare the current customer's post-loan rule model score result with the risk decision threshold pre-configured according to the customer group's differentiation, and determine and execute post-loan risk control decisions based on the comparison results.

7. The strategy-based hierarchical risk control method based on an integrated model according to claim 6, characterized in that, The post-loan integrated model is constructed and trained to obtain customer rating indicator features; The process of predicting customer default probabilities using a trained post-loan ensemble model and integrating the prediction results to obtain the default probabilities for all customers includes the following steps: Based on customer rating indicators and combined with the Light-GBM algorithm, a post-loan integrated model is constructed. The hyperparameters of the post-loan ensemble model are optimized using grid search to obtain the trained post-loan ensemble model. Obtain customer rating indicator features and, based on the trained post-loan ensemble model, obtain the customer's post-loan ensemble model score result; According to the preset interval rules, the customer's post-loan integrated model scoring results are divided into several bins, and the bins are mapped in combination with the penalty coefficient to obtain the customer's default probability. By integrating the prediction results, the default probability of rejecting customers and the default probability of passing customers are obtained separately.

8. The strategy-based hierarchical risk control method based on an integrated model according to claim 6, characterized in that, The process of using a genetic algorithm to optimize the hard rules and determine the final hard rules includes the following steps: A fitness function is constructed using pass rate, default rate, and profit margin. Based on the fitness function, the fitness value of the hard rule is calculated. By using affinity crossover functions, we can perform affinity crossover and mutation on hard rules to obtain the offspring of hard rules. Based on the fitness value of the hard rule, the offspring of the hard rule are selected and the best are retained. The next generation subset is generated by combining roulette wheel and the final hard rule is determined through iterative optimization.

9. The strategy-based hierarchical risk control method based on an integrated model according to claim 6, characterized in that, The step of comparing the current customer's post-loan rule model scoring result with the risk decision threshold pre-configured based on customer group differentiation, and determining and executing post-loan risk control decisions based on the comparison results, includes: Risk decision thresholds are configured according to customer group differences and divided into a first risk decision threshold and a second risk decision threshold, wherein the first risk decision threshold is greater than the second risk decision threshold. If the current customer’s post-loan rule model score is greater than or equal to the first risk decision threshold, determine and execute the first post-loan preset risk control decision. If the current customer's post-loan rule model score is greater than or equal to the second risk decision threshold and less than the first risk decision threshold, then determine and execute the second post-loan preset risk control decision.

10. A strategy-based hierarchical risk control system based on an integrated model, characterized in that, The system includes: The post-loan risk prediction module is used to build and train the post-loan integrated model to obtain customer rating indicator characteristics; combine the trained post-loan integrated model to predict the customer's default probability, and integrate the prediction results to obtain the default probability of all customers. The post-loan preset risk control rule classification module is used to calculate the effectiveness of preset risk control based on the default probability of all customers, and to classify preset risk control rules into soft rules and hard rules in combination with the grading standard. The post-loan rule scoring model construction module is used to optimize the hard rules using a genetic algorithm to determine the final hard rules; and to construct the post-loan rule scoring model based on the final hard rules and soft rules. The post-loan strategy tiered risk control decision module is used to obtain the rating indicator characteristics of the current customer, calculate the post-loan rule scoring model using the post-loan rule scoring model, obtain the post-loan rule model scoring result of the current customer, compare the post-loan rule model scoring result of the current customer with the risk decision threshold pre-configured according to the customer group differentiation, and determine and execute the post-loan risk control decision based on the comparison result.

Images