Bad loan collection method

Abstract

The invention relates to the technical field of data processing, and discloses a bad loan collection method, which comprises the following steps: data acquisition and preprocessing: obtaining loan related information, repayment behavior data and overdue records of customers, and carrying out data cleaning and feature extraction; customer risk modeling and layering are carried out, risk scores are generated based on customer behavior data, and risk layering management is carried out on customers; and generating and optimizing a collection strategy, generating a collection strategy matrix according to the customer risk category and the overdue state, and selecting an optimal collection mode through a dynamic optimization model. According to the method, the logistic regression and the time sequence model are combined to realize the fusion evaluation of static and dynamic behaviors, and the accurate prediction of the customer repayment capability is improved; the collection mode is dynamically optimized through the collection strategy matrix, and the resource utilization rate is increased; a closed-loop feedback mechanism is adopted, and the optimization real-time performance is enhanced; and a multi-dimensional index is introduced to construct a reward function, the rate of return and the collection cost are comprehensively considered, and the collection income and economy are optimized.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and specifically to a method for collecting overdue loans. Background Art

[0002] Debt collection is an important part of the entire credit risk management process. It can not only effectively reduce the asset losses of banks, but also beneficially feedback on customer marketing work, loan granting policies, model optimization, and customer relationship management; With the development of personal consumer loan business, especially the development of personal online loans and personal online loans in recent years, the personal loan business has developed towards convenience, small amounts, and multiple transactions. Intelligent monitoring and management of overdue cases generated by banking business can achieve integrated management of debt collection in post-loan risk management work; Generally speaking, the existing technology mainly uses logistic regression or decision tree models to statically evaluate customer risks. This method can simply analyze the basic information and historical repayment data of customers to obtain their risk scores. However, this static scoring model can only reflect the historical status of customers and is difficult to dynamically capture the real-time changes in customer behavior. For example, when a customer shows a positive willingness to repay during the debt collection process, these models cannot quickly adjust the score, resulting in lagging or ineffective strategies; in addition, traditional debt collection strategies often rely on a fixed rule matrix. These rules usually divide debt collection methods according to the customer's risk level and overdue days, such as SMS reminders, phone calls, etc. Although these rules consider the conventional customer stratification in the design, they lack the ability to dynamically adjust. When the overdue days of a customer increase or the repayment willingness changes, the fixed rules cannot flexibly respond, easily causing waste of debt collection resources or a decline in debt collection effects; In the existing technology, the optimization of debt collection strategies usually relies on manual experience or simple statistical analysis. This method is not only inefficient, but also easily interfered by subjective factors of humans and lacks scientificity. For example, when the debt collection response rate of a customer is low, it is difficult for the existing technology to clearly judge whether it is a problem with the debt collection method or the customer's own low willingness through quantitative methods. This ambiguity limits the improvement effect of the strategy; More importantly, the existing technology has a low utilization rate of the result data after the execution of debt collection tasks. Many debt collection systems only record the data after the task is completed without further feedback and analysis. This one-way process ignores the verification of the effectiveness of debt collection behaviors and fails to optimize subsequent strategies through result data, resulting in a lack of closed-loop and continuous improvement capabilities in the entire debt collection process. Therefore, a method for collecting overdue loans is proposed to solve the above problems. Summary of the Invention

[0003] In view of the deficiencies of the prior art, the present invention provides a non-performing loan collection method, which solves the problems that due to the use of a static scoring model and a fixed rule matrix, the changes in customer behavior cannot be dynamically reflected, resulting in lagging strategies and poor effects. At the same time, the optimization of collection strategies relies on manual experience, lacks scientificity, and the task execution results do not form a closed-loop feedback, resulting in difficulties in continuously improving the collection process and low resource utilization rate.

[0004] To achieve the above objectives, the present invention is realized through the following technical solutions: A non-performing loan collection method includes the following steps: Data collection and preprocessing, obtaining the loan-related information, repayment behavior data, and overdue records of customers, and performing data cleaning and feature extraction; Customer risk modeling and stratification, generating a risk score based on customer behavior data, and performing risk stratification management on customers; Collection strategy generation and optimization, generating a collection strategy matrix according to the customer risk category and overdue status, and selecting the optimal collection method through a dynamic optimization model; Collection task execution, allocating collection tasks according to the collection strategy, and recording collection behaviors and customer repayment behaviors; Closed-loop feedback, feeding back the results of collection behaviors to the system, updating customer scores, and optimizing collection strategies.

[0005] Preferably, the data collection and preprocessing include: Obtaining the loan amount, overdue days, and historical repayment ratio of customers; Filling missing values with the mean value, and removing outliers using the standard deviation method; Normalizing the data to make the numerical features meet the modeling requirements.

[0006] Preferably, the customer risk modeling and stratification include the following steps: Calculating the overdue probability of customers through a logistic regression model, with the input including the loan amount, overdue days, and historical repayment behavior; using a time series model to capture the dynamic changes in customer behavior, and generating a risk score based on historical behavior; Dividing customers into high-risk, medium-risk, and low-risk categories according to the customer risk score.

[0007] Preferably, the time series model is a deep learning model based on LSTM, using the customer behavior sequence as the input to generate the risk score of the time step, and the hidden state is updated through the following formula: h t =σ(W h h t-1 +W x X t,3 +b h ) Where ht is the hidden state of the LSTM model at time step t, representing the internal representation of the customer risk at the current moment; h t-1 is the hidden state of the LSTM model at time step t - 1; W h is the hidden state weight matrix; W x is the input layer weight matrix; X t，3 is the input feature at time step t; b h is the bias term; σ(·) is the activation function.

[0008] Preferably, the collection strategy generation and optimization include: Generate a collection strategy matrix based on the customer risk category and overdue days. The matrix includes collection methods, execution entities, and execution frequencies; use a reinforcement learning model to dynamically optimize the strategy matrix, where the state includes the customer risk category and overdue days, the action is the collection method, and the reward value is calculated from the customer repayment amount and collection cost.

[0009] Preferably, the calculation formula for the reward value is R t = M recovered - C action where, R t is the reward value, representing the return of the collection behavior; M recovered is the repayment amount of the collection behavior; C action is the cost of the collection behavior.

[0010] Preferably, the collection strategy matrix includes the following: SMS collection, used for the early overdue stage of low-risk customers; Phone collection, used for the mid-term overdue stage of medium-risk customers; Outsourced collection, used for cases of high-risk customers with overdue days exceeding 90 days; Legal collection, used for customers with a relatively high risk score and ineffective collection methods.

[0011] Preferably, the collection task execution includes: Allocate collection tasks according to the customer risk score and the strategy matrix. The task priority is the product of the risk score and the task frequency; high-priority tasks are executed by manual collection and the legal department, and low-priority tasks are completed by SMS collection and robot collection.

[0012] Preferably, the closed-loop feedback includes the following steps: Record the customer's repayment behavior and the results of the collection behavior; Update the customer risk score and input the latest behavior data into the risk modeling module; Optimize the collection strategy matrix, and use the collection recovery rate and behavioral cost as the input of the reinforcement learning reward function.

[0013] The present invention also provides a non-performing loan collection system, including: A data collection and preprocessing module, which is used to collect the loan information and behavioral data of customers and perform feature construction and normalization; A risk modeling module, which is used to calculate the risk score of customers and classify risk categories; A strategy generation module, which is used to generate a collection strategy matrix and dynamically optimize it; A collection task module, which is used to assign tasks according to the strategy and record the results of collection behaviors; A feedback module, which is used to feedback the results of collection behaviors to the risk model and the strategy matrix to complete the optimization closed-loop.

[0014] The present invention provides a non-performing loan collection method. It has the following beneficial effects: 1. The technical solution of the present invention combines a logistic regression model and a time series model (LSTM), realizing the effective integration of static risk assessment and dynamic behavior prediction. Through this combined model, it can accurately capture the static characteristics and dynamic behavior changes of customers. Compared with the technical solution in the prior art where a single static model is difficult to adapt to real-time behavior changes, it solves the problem that dynamic data cannot timely reflect the true repayment ability of customers.

[0015] 2. By designing a collection strategy matrix and introducing a reinforcement learning model for dynamic optimization, the present invention achieves the technical effect of real-time adjustment of collection strategies. This technical solution can flexibly select collection methods according to the risk changes and overdue status of customers. Compared with the deficiencies in the prior art where collection strategies are rigid and cannot be dynamically adapted to customer behaviors, the present invention significantly improves the utilization rate of collection resources and the accuracy of strategy matching.

[0016] 3. The present invention adopts a closed-loop feedback mechanism, and updates the results of collection task execution to the risk model and the strategy matrix in real time, thus forming an automated optimization cycle. By collecting and analyzing collection data in real time, it dynamically optimizes the assignment of collection tasks. Compared with the problems of low efficiency and strategy lag caused by manually adjusting collection strategies in the prior art, the present invention effectively solves the problem of insufficient real-time performance of strategy optimization.

[0017] 4. By introducing multi-dimensional indicators such as the recovery rate and collection cost, constructing a reward function, and combining reinforcement learning to optimize the collection strategy in a revenue-driven manner, the present invention achieves the technical goal of improving the collection effect while reducing the collection cost. Compared with the limitations of measuring collection effects with a single indicator in the prior art, the present invention comprehensively considers economy and profitability, and solves the problems of imbalance and simplification of collection schemes. Brief Description of the Drawings

[0018] Figure 1 It is a schematic flow chart of the method of the present invention; Figure 2 It is a schematic diagram of the system framework of the present invention; Detailed Embodiments

[0019] Next, in combination with the attached drawings of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0020] Please refer to the attached Figure 1 , the embodiments of the present invention provide a non-performing loan collection method, including the following steps: S1. Data collection and preprocessing, obtaining loan-related information, repayment behavior data, and overdue records of customers, and performing data cleaning and feature extraction; Specifically, this step S1 aims to provide high-quality basic data support for subsequent risk modeling, strategy generation, and optimization modules in non-performing loan collection. The data collection and preprocessing link is the core pre-step of the entire collection method. It mainly collects, cleans, constructs features, and standardizes the relevant information of customers to ensure the accuracy and applicability of the input data for the subsequent model. Specifically, the results of data processing will directly serve as the input for the risk modeling module, and its influence runs through the entire execution process of the collection method. Therefore, in technical implementation, it is necessary to ensure the comprehensiveness and scientificity of data processing.

[0021] In this embodiment, the content of data collection includes but is not limited to the following types: Basic information data of customers, such as loan amount, repayment history record, overdue days, etc.

[0022] Behavior data, such as the repayment willingness of customers and the historical repayment ratio.

[0023] External credit information data, such as the mortgage situation and value of collateral.

[0024] In some embodiments, the consumption behavior data of customers can also be collected to further enrich the input required for modeling.

[0025] Specifically, in the data collection link, generally, the above data content can be obtained through internal data sources (such as the bank's internal management system) and external data sources (such as third-party credit investigation systems). As an option, in certain scenarios, real-time dynamic data can be introduced, such as the current overdue status of customers or the time of the most recent repayment.

[0026] After data collection is completed, data cleaning is required. Data cleaning includes, but is not limited to, the following: Fill in missing values. Generally, the mean value filling method can be used to process numerical data, while the mode filling method can be used for categorical data.

[0027] Remove outliers. As an implementation method, the method based on three times the standard deviation can be used to process abnormal data.

[0028] Delete duplicate records to ensure data consistency and accuracy.

[0029] In some embodiments, data normalization processing can be introduced. The goal of normalization processing is to eliminate the influence of the eigenvalue range on the model training results. Specifically, the following formula is used for normalization processing in this embodiment: ′ where X i，1 represents the normalized eigenvalue; X i，1 represents the original value of the i-th feature; max(X 1 ) represents the maximum value of the eigenvalue; min(X 1 ) represents the minimum value of the eigenvalue.

[0030] In the process of feature construction, the following key features are extracted in this embodiment: Overdue days, defined as the cumulative overdue days from the agreed repayment date to the current time of the customer.

[0031] Repayment ratio, the calculation formula is: where P paid represents the repayment ratio; M paid represents the amount of money the customer has repaid; M loan represents the total loan amount.

[0032] Mortgage rate, defined as the ratio of the value of the customer's collateral to the loan amount. Specifically, the calculation formula of the mortgage rate is: where R pledge is the mortgage rate; V collateral is the value of the collateral; M loan is the loan amount.

[0033] In a possible implementation method, whether the customer has collateral can also be used as a binary variable for feature construction, with a value of "1" indicating the existence of collateral and "0" indicating the non-existence.

[0034] Generally, in the feature construction stage, a feature selection strategy can be introduced. For example, the method of feature importance evaluation can be used to screen out the variables that have a greater impact on the modeling results, so as to reduce the interference of redundant features. In some embodiments, in combination with business rules, core variables such as collateral value and overdue days can be preferentially used as model inputs.

[0035] S2. Customer risk modeling and stratification, generating a risk score based on customer behavior data, and conducting risk stratification management for customers; Specifically, customer risk modeling and stratification is one of the key steps of the present invention. The main objective is to evaluate the repayment ability and default risk of customers based on the high-quality data processed in the previous step S1. By establishing an accurate risk scoring model, it provides a reliable reference basis for the generation of subsequent collection strategies. This step depends on both the global analysis of historical data and the dynamic changes in customer behavior. The output of the model is directly used for customer risk stratification, which affects the selection of collection methods.

[0036] In this embodiment, the process of risk modeling includes two parts: static risk assessment and dynamic behavior prediction. Generally, static risk assessment can be performed through a logistic regression model. The logistic regression model is a commonly used linear classifier that can effectively handle classification tasks. As an option, the default probability of a customer can be calculated by the following formula: Specifically, P(y = 1|X 2 ) represents the probability that a customer is overdue, β 0 is the intercept term of the model, β i is the regression coefficient of the i-th feature, and X i，2 is the input value of the i-th feature. The inputs of the logistic regression model include but are not limited to loan amount, overdue days, and historical repayment ratio.

[0037] In a possible implementation, in order to further improve the prediction accuracy of the model, a time series model can be introduced to capture the dynamic changes in customer behavior. For example, the long short-term memory network (LSTM) is used to model the historical behavior of customers to generate risk scores for consecutive time steps. The LSTM model has a memory ability and can effectively process time series data. Generally, the hidden state of the LSTM can be updated by the following formula: h t = σ(W h h t-1 + W x X t,3 + b h ) where h t$h_t$ is the hidden state of the LSTM model at time step $t$, representing the internal representation of the customer risk at the current moment; t-1 $h_{t - 1}$ is the hidden state of the LSTM model at time step $t - 1$; h $W_h$ is the hidden state weight matrix; x $W_x$ is the input layer weight matrix; t，3 $x_t$ is the input feature at time step $t$; h $b$ is the bias term; $\sigma(\cdot)$ is the activation function.

[0038] As an implementation, the input features of the LSTM model can include the number of overdue days, the proportion of repaid amount, and the historical response behavior of the customer. Specifically, in some scenarios, the behavior response time can be introduced as an input feature of the time series to capture the feedback pattern of the customer to the collection method. For example, when the customer makes a positive response within a short time, this feature can enhance the dynamic adjustment ability of the risk score.

[0039] In some embodiments, a combined model can be adopted to combine the static risk score of logistic regression with the dynamic risk score generated by the LSTM. The combined model can consider both the static impact of historical behavior and the dynamic changes of real-time behavior. In this combined model, the final risk score can be expressed as the following formula: $R$ final $=$ 1 $\alpha$ static $\cdot R_s$ 1 $+(1 - \alpha)$ dynamic where $R$ final represents the final risk score, $R_s$ static is the static risk score generated by the logistic regression model, $R_d$ dynamic is the dynamic risk score generated by the LSTM, and $\alpha$ 1 is the weight factor of the static score, with a value range of $[0, 1]$.

[0040] Specifically, the weight factor of the combined model can be optimized according to historical data. In one possible implementation, the optimal $\alpha$ 1 value can be found through the grid search method to maximize the prediction accuracy of the model.

[0041] In this embodiment, the output results of the customer risk score will be hierarchically managed according to a preset threshold. Generally, the risk score can be divided into high risk ($R$ final $> 0.8$), medium risk ($0.5 \leq R$ final $\leq 0.8$), and low risk ($R$ final $< 0.5$) categories. As an option, these thresholds can be adjusted according to the actual effect of the collection strategy.

[0042] In some embodiments, additional rules can be introduced to correct the stratification results. For example, for customers with a relatively high mortgage rate, their risk stratification can be preferentially reduced; while for customers with an abnormally high number of historical overdue days, their risk level can be increased. This correction mechanism can further improve the accuracy of risk stratification and provide stronger guarantee for the precise implementation of subsequent collection strategies.

[0043] S3. Collection strategy generation and optimization: Generate a collection strategy matrix based on the customer risk category and overdue status, and select the optimal collection method through a dynamic optimization model; Specifically, collection strategy generation and optimization is a key step in formulating personalized collection methods for customers of different risk categories and their overdue statuses based on customer risk modeling and stratification. In this step, by generating a collection strategy matrix, precise matching between customers and collection means is achieved, and at the same time, the strategy is dynamically optimized with the help of a reinforcement learning model to make the allocation of collection resources more efficient. The output of the collection strategy matrix directly affects the execution of subsequent collection tasks, and its optimization results are closely related to customer repayment rates, collection costs, etc.

[0044] In this embodiment, the generation and optimization of the collection strategy matrix mainly include two parts: one is to generate a collection strategy matrix, defining the collection methods, execution entities, and frequencies for customers of different risk categories; the other is to dynamically optimize the strategy through a reinforcement learning model so that the collection methods can adapt to customer behavior changes and maximize the repayment income.

[0045] In the process of generating the strategy matrix, generally, the core rules of the matrix can be constructed based on the customer's risk score and overdue days. As an option, each entry of the matrix can be defined as a combination of a collection method, an execution entity, and an execution frequency. For example, for the early overdue stage of low-risk customers, use the method of sending text message reminders once every 3 days; while for the late overdue stage of high-risk customers, use the method of legal litigation once every 15 days. Specifically, the collection strategy matrix can be expressed as the following set of rules: In a possible implementation, the above rules can be optimized based on historical data. For example, by statistically analyzing the repayment rates of collection methods under different overdue days, the execution frequency and the selection of collection methods can be dynamically adjusted.

[0046] To further improve the adaptability of the collection strategy, a reinforcement learning model is introduced in this embodiment to dynamically optimize the collection strategy matrix. Generally, the core of the reinforcement learning model lies in defining the state, action, and reward function. Specifically, the states defined in this embodiment include the customer's risk level, overdue days, and historical repayment ratio; the actions are the selection of collection methods, such as SMS collection, phone collection, legal litigation, etc.; the reward value is calculated based on the customer's repayment amount and collection cost, and its formula is as follows: R t =M recovered -C action Wherein, R t is the reward value, representing the income of the collection behavior; M recovered is the repayment amount of the collection behavior; C action is the cost of the collection behavior.

[0047] In a possible implementation, the reinforcement learning model uses the Q-Learning algorithm to update the policy matrix, and its value function update formula is: Q(S t ,A t )=Q(S t ,A t )+α 2 [R t +γmax A′ Q(S t+1 ,A ′ )-Q(S t ,A t )] Wherein, Q(S t ,A t ) represents the value of selecting action A t in state S t ; α 2 represents the learning rate, controlling the learning speed of the model for new information; R t is the reward value; γ is the discount factor, controlling the influence weight of future rewards; max A′ Q(S t+1 ,A ′ ) represents the optimal action value in the next state S t+1 .

[0048] As an implementation, the parameters of the reward function can be adjusted according to the actual collection effect. For example, in a scenario where the repayment rate is emphasized, the weight of M recovered can be increased; while in a scenario where cost control is prioritized, the weight of C action can be increased.

[0049] In some embodiments, the initial policy of the reinforcement learning model can also be set to accelerate the training process of the model. For example, the collection method with the highest repayment rate in historical data can be used as the initial policy, and the policy matrix can be gradually optimized through actual feedback.

[0050] S4. Execution of collection tasks: According to the collection policy, assign collection tasks and record collection behaviors and customer repayment behaviors. Specifically, the execution of collection tasks is one of the core links in the implementation of the present invention. Its purpose is to assign specific collection tasks to appropriate execution entities based on the collection policy matrix generated and optimized in step S3, and effectively record the collection results. The efficiency and accuracy of the task execution link directly affect the overall collection effect, and at the same time provide key input data for closed-loop feedback. This step comprehensively evaluates the risk score of the customer and the policy priority to ensure the reasonable allocation and optimal utilization of collection resources.

[0051] In this embodiment, the generation and execution of collection tasks mainly involve the following aspects: determination of task priority, formulation of task assignment rules, and recording of collection behaviors.

[0052] Generally, the priority of collection tasks is calculated based on the risk score of the customer and the frequency of collection tasks. Specifically, the priority calculation formula is as follows: T a = S a × F a Wherein, T a represents the task priority of customer a, and the larger the value, the higher the priority; S a represents the risk score of customer a, provided by the risk modeling module; F a represents the frequency of collection tasks for customer a, determined by the policy matrix. As an option, the task priority can also be dynamically adjusted in combination with the historical repayment behavior of the customer. For example, for customers with a relatively recent last repayment time, the priority can be appropriately reduced to avoid over-collection.

[0053] In a possible implementation manner, the task assignment rule of collection tasks can be hierarchically assigned according to the customer risk level and priority value. Specifically: High-priority tasks are generally assigned to manual collection personnel or the legal department for execution, and are applicable to high-risk customers and long-term overdue cases.

[0054] Medium-priority tasks can be carried out by telephone collection and are applicable to medium-risk customers and the medium-term overdue stage.

[0055] Low-priority tasks are usually completed by SMS collection or robot collection and are applicable to low-risk customers and the early overdue stage.

[0056] In some embodiments, restrictions on task allocation can also be introduced. For example, when the workload of a certain debt collection entity exceeds a preset threshold, the task can be reallocated to avoid overuse or concentration of resources.

[0057] In this embodiment, the records of debt collection behaviors include the following: call records, text message replies, and repayment amounts, etc. Generally, call records can be used to analyze the customer's response degree to debt collection, while text message replies can reflect the customer's repayment intention. As an option, the record of repayment amount includes not only the actual recovered amount, but also the customer's installment repayment plan.

[0058] S5. Closed-loop feedback, feedback the results of debt collection behaviors to the system, update the customer score and optimize the debt collection strategy.

[0059] Specifically, closed-loop feedback is a crucial step in the technical solution of the present invention. It aims to dynamically update the customer risk score and the debt collection strategy matrix by collecting and analyzing the execution results of debt collection tasks, so as to form a complete feedback mechanism. This step takes the execution results of debt collection tasks as input, combines with the real-time recorded data, optimizes the model and strategy, and ensures that the debt collection process can be continuously improved. Closed-loop feedback directly affects the effect of subsequent task execution and runs through the entire cycle of the debt collection method.

[0060] In this embodiment, the implementation of closed-loop feedback includes three core parts: data recording, result analysis, and strategy optimization.

[0061] Generally, the result data of debt collection behaviors need to be recorded completely and in a timely manner. As an option, the result data can include the customer's repayment amount, repayment time, repayment method, and the customer's behavioral response. For example, for phone debt collection, it can be recorded whether the customer answers the phone and the response attitude towards the debt collection content; while for text message debt collection, it can be recorded whether the customer replies and the intention of the reply content.

[0062] Specifically, the customer's repayment behavior data will be stored in a structured form, for example: Customer repayment amount, which is used to quantify the actual effect of debt collection tasks.

[0063] Customer repayment time, which is used to evaluate the timeliness of different debt collection methods.

[0064] Customer's installment repayment plan, which is used to track the customer's future repayment behavior.

[0065] In a possible implementation manner, natural language processing (NLP) technology can also be introduced to perform semantic analysis on the customer's text replies. For example, label the reply content as "promise to repay" or "refuse to repay", and adjust the customer's risk score and debt collection strategy according to the analysis results.

[0066] In this embodiment, the result data will be analyzed through an algorithm model to evaluate the effectiveness of different collection methods and the changing trends of customer behavior. For example, the recovery rate of different collection methods can be calculated, which is expressed by the following formula: Where, R payback represents the recovery rate; M recovered represents the actual recovered amount; M due represents the due receivable amount.

[0067] In terms of strategy optimization, in this embodiment, a reinforcement learning model is adopted to dynamically adjust the collection strategy matrix. Specifically, the reward function takes the recovery rate and collection cost in the collection result as inputs. As a possible implementation, the reward function can be defined as: R t = w 1 ·R payback - w 2 ·C action Where, R t represents the reward value; w 1 and w 2 are the weight factors of the recovery rate and collection cost respectively; R payback represents the recovery rate; C action represents the execution cost of the collection task, including telephone charges, text message charges or legal litigation charges.

[0068] In a possible implementation, the weight factors w 1 and w 2 can be optimized through regression analysis of historical data to ensure that the reward function can better reflect the real benefits of collection behavior.

[0069] In this embodiment, the results of strategy optimization will be fed back to the risk modeling module and the collection task execution module in real time. For example, when the effectiveness of a certain collection method is significantly improved, its weight can be increased so that it can be preferentially adopted in subsequent tasks.

[0070] As an extension, in some embodiments, a long-term feedback analysis mechanism can also be introduced. For example, on a monthly or quarterly basis, the results of all collection behaviors are statistically summarized, and accordingly, the customer stratification rules and the generation rules of the collection strategy matrix are adjusted.

[0071] Please refer to the appendix Figure 2 , the present invention also provides a non-performing loan collection system, including: The data collection and preprocessing module is used to collect the loan information and behavioral data of customers and perform feature construction and normalization. Specifically, this module is the input end of the system, responsible for collecting the basic information and behavioral data of customers and normalizing the data.

[0072] Generally, the data collection sources include internal databases (such as loan systems and customer management systems) and external data sources (such as credit bureaus). The collected content covers static features such as the loan amount, overdue days, historical repayment records, and guarantee information of customers, as well as dynamic features such as the most recent repayment behavior and overdue status.

[0073] In terms of feature construction, key features such as overdue days, repayment ratio, and mortgage rate can be extracted, and the data can be further enriched through combined features. For example, the ratio of the mortgage value of the collateral to the loan amount is used as an important indicator to help the model more accurately evaluate the repayment ability of customers.

[0074] The risk modeling module is used to calculate the risk score of customers and classify the risk categories. The risk modeling module is responsible for generating the risk score of customers through data analysis and modeling methods and classifying customers into different risk levels.

[0075] Generally, this module combines two methods: static modeling and dynamic modeling. Static modeling is mainly based on the logistic regression model, and the overdue probability is calculated by analyzing the basic characteristics of customers. Dynamic modeling uses time series models (such as LSTM) to capture the changing trends of customer behavior data and generate the dynamic risk score of customers.

[0076] As a possible implementation method, the risk modeling results will be output through the method of multi-model fusion. For example, the static score and the dynamic score are combined according to certain weights to obtain the comprehensive risk score. Customers are classified into high-risk, medium-risk, and low-risk categories according to the score value, which is used to guide the generation of subsequent collection strategies.

[0077] The strategy generation module is used to generate the collection strategy matrix and dynamically optimize it. This module generates the collection strategy matrix based on the risk score and overdue status of customers and improves the adaptability of the strategy through dynamic optimization.

[0078] Generally, the collection strategy matrix defines the collection methods, execution entities, and execution frequencies for customers in different risk categories. As an option, SMS collection can be used for low-risk customers, and mandatory measures such as legal litigation are preferred for high-risk customers.

[0079] The matrix generated by the strategy will be further optimized through historical collection data. For example, the strategy matrix is adjusted through a reinforcement learning model so that the collection methods can dynamically adapt to changes in customer behavior. The goal of optimization is to balance collection costs and repayment benefits and improve the overall collection efficiency.

[0080] A collection task module, used to allocate tasks according to the strategy and record the results of collection behaviors; The collection task module generates specific collection tasks according to the strategy matrix and allocates them to the corresponding execution entities.

[0081] When generating tasks, the risk score, collection frequency, and historical behavior of the customer are usually combined. High-priority tasks are assigned to human collectors or legal teams, and low-priority tasks are completed by SMS or call bots. During the allocation process, the tasks are evenly adjusted according to the workload of the execution entity to avoid resource overload.

[0082] In some embodiments, the system monitors the collection process. For example, through speech recognition technology, the call content of telephone collections is analyzed in real time to ensure the compliance of collection behaviors.

[0083] At the same time, the module records all collection behaviors, including whether the customer repays, the repayment amount, the repayment time, and the customer's response to the collection method. These records will be directly input into the feedback module to provide data support for strategy optimization.

[0084] A feedback module, used to feedback the results of collection behaviors to the risk model and the strategy matrix to complete the optimization closed-loop.

[0085] The feedback module is the core of the system's closed-loop process, responsible for analyzing the execution results of collection tasks and optimizing the strategy and risk scoring model accordingly.

[0086] Generally, the feedback module will collect data including the customer's repayment amount, repayment time, installment repayment plan, and the customer's response attitude. Through these data, the system can evaluate the effectiveness of different collection methods and dynamically adjust the strategy matrix. For example, if a certain collection method is effective in a certain risk level, its weight in the strategy matrix can be increased.

[0087] In some scenarios, semantic analysis can also be performed on the customer's text or voice feedback. For example, the customer's reply is labeled as "committed to repayment" or "refused to repay", and the result is feedback to the risk model to adjust the customer's risk score in real time.

[0088] In terms of strategy optimization, the system retrains the reinforcement learning model based on the collection results and historical data. The optimized strategy is directly used for the next round of task allocation to achieve adaptive adjustment.

[0089] Although embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for collecting bad loans, characterized in that: The following steps are involved: Data collection and preprocessing: obtaining customers’ loan-related information, repayment behavior data, and overdue records, and performing data cleaning and feature extraction; Customer risk modeling and stratification: generating risk scores based on customer behavior data and conducting risk stratification management for customers; Collection strategy generation and optimization: Generate a collection strategy matrix based on customer risk category and overdue status, and select the optimal collection method through a dynamic optimization model; Execute collection tasks, assign collection tasks according to collection strategies, and record collection behaviors and customer repayment behaviors; Closed-loop feedback feeds collection results back to the system, updates customer ratings and optimizes collection strategies.

2. A method for collecting bad loans according to claim 1, characterized in that: The data collection and preprocessing include: Obtain the customer's loan amount, overdue days and historical repayment ratio; Missing values ​​were filled with mean values, and outliers were removed using the standard deviation method; Normalize the data so that the numerical features meet the modeling requirements.

3. A method for collecting bad loans according to claim 1, characterized in that: The customer risk modeling and stratification includes the following steps: The customer's probability of default is calculated using a logistic regression model, with inputs including loan amount, number of days overdue, and historical repayment behavior; Use time series models to capture dynamic changes in customer behavior and generate risk scores based on historical behavior; Customers are classified into high risk, medium risk and low risk categories based on their risk scores.

4. A method for collecting bad loans according to claim 3, characterized in that: The time series model is a deep learning model based on LSTM, which uses the customer behavior sequence as input to generate the risk score of the time step, and the hidden state is updated by the following formula: h t =σ(W h h t-1 +W x X t +b h ) Among them, h t is the hidden state of the LSTM model at time step t, indicating the internal representation of customer risk at the current moment; h t-1 is the hidden state of the LSTM model at time step t-1; W h is the hidden state weight matrix; W x is the input layer weight matrix; X t，3 is the input feature at time step t; b h is the bias term; σ(·) is the activation function.

5. A method for collecting bad loans according to claim 1, characterized in that: The collection strategy generation and optimization includes: A collection strategy matrix is ​​generated based on customer risk categories and overdue days. The matrix includes collection methods, execution entities, and execution frequencies. A reinforcement learning model is used to dynamically optimize the strategy matrix, in which the states include customer risk categories and overdue days, the actions are collection methods, and the reward value is calculated based on the customer repayment amount and the collection cost.

6. A method for collecting bad loans according to claim 5, characterized in that: The calculation formula of the reward value is: R t =M recovered -C action Among them, R t is the reward value, indicating the benefit of the collection behavior; M recovered The amount of money collected from the collection action; C action The cost of debt collection activities.

7. A method for collecting bad loans according to claim 1, characterized in that: The collection strategy matrix includes the following: SMS collection is used for low-risk customers in the early stages of overdue payments; Telephone collection is used for medium-risk customers in the mid-term overdue stage; Outsourcing debt collection is used for cases involving high-risk customers with overdue days exceeding 90 days; Legal collection is used for customers with higher risk scores and ineffective collection methods.

8. A method for collecting bad loans according to claim 1, characterized in that: The collection task execution includes: Assign collection tasks based on customer risk scores and strategy matrix, with the task priority being the product of risk score and task frequency; High-priority tasks are performed through manual collection and the legal department, while low-priority tasks are completed through SMS collection and robot collection. The execution results of collection tasks are recorded, including call records, SMS replies, and repayment amounts.

9. A method for collecting bad loans according to claim 1, characterized in that: The closed-loop feedback comprises the following steps: Record the customer's repayment behavior and collection results; Update customer risk scores and input the latest behavioral data into the risk modeling module; Optimize the collection strategy matrix and use the collection rate and behavior cost as the input of the reinforcement learning reward function.

10. A system for collecting bad loans, applied to a method for collecting bad loans as claimed in any one of claims 1 to 9, characterized in that: include: Data collection and preprocessing module, used to collect customers' loan information and behavior data and perform feature construction and normalization; Risk modeling module, which is used to calculate the risk score of customers and classify risk categories; Strategy generation module, used to generate the collection strategy matrix and dynamically optimize it; The collection task module is used to assign tasks according to strategies and record the results of collection actions; The feedback module is used to feed back the results of collection behavior to the risk model and strategy matrix to complete the optimization closed loop.

Images