A method, device and equipment for dynamic assessment of post-loan risk and a storage medium

Abstract

The embodiment of the application relates to the technical field of post-loan risk management, in particular to a post-loan risk dynamic evaluation method, device and equipment and a storage medium. The method comprises the following steps: collecting real-time overdue data of a client; the overdue data comprises client information, loan information, current overdue days, current repayment status and sample quantity; based on the overdue data, a multi-dimensional dynamic weight matched with the current overdue state of the client is calculated; the multi-dimensional dynamic weight comprises overdue day importance weight and sample weight; the multi-dimensional dynamic weight is fused and normalized to obtain a normalized weight; the normalized weight and the overdue data are input into a pre-deployed risk prediction model for inference processing to generate a real-time risk score of the client, so that a corresponding collection strategy is generated according to the real-time risk score of the client. The method can realize real-time dynamic and accurate evaluation of overdue user risks.

Description

Technical Field

[0001] This application relates to the field of post-loan risk management technology, and in particular to a method, apparatus, device and storage medium for dynamic assessment of post-loan risk. Background Technology

[0002] In the field of post-loan risk management, accurately identifying the repayment willingness and risk level of overdue users is a core requirement for optimizing collection strategies and improving collection efficiency. As a key tool for risk assessment, the post-loan scoring card model's core function is to rank overdue users based on their characteristics, matching differentiated collection resources to users with different risk levels.

[0003] Traditional scoring card models generally adopt a static point-in-time scoring mode, which selects fixed milestone time points after delinquency (such as the 1st day or the 7th day of delinquency) to calculate risk scores. This mode has gradually exposed many technical defects in long-term application: First, static point-in-time scoring can only capture the risk status of users at specific time points and cannot track the dynamic changes in users' risk performance as the number of delinquent days increases, resulting in risk assessment lagging behind the actual situation; Second, it lacks an effective knowledge transfer mechanism and is difficult to take into account the risk characteristics of customers with long and short delinquencies; Third, it ignores the heterogeneity of sample distribution at different delinquency stages.

[0004] Therefore, how to achieve real-time, dynamic, and accurate assessment of the risks of overdue users is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] The purpose of this application is to provide at least one method, device, equipment, and storage medium for dynamic post-loan risk assessment, which can realize the transformation from static point-in-time scoring to real-time dynamic scoring, effectively capture the user's latest repayment intention, and at the same time balance the training weights of long-term and short-term overdue samples through a multi-dimensional dynamic weighting mechanism, improve the generalization ability and accuracy of the risk assessment model, and provide reliable support for the formulation of differentiated collection strategies.

[0006] To address the aforementioned technical problems, at least one embodiment of this application provides a method for dynamic post-loan risk assessment, comprising: Collect real-time overdue data from customers; the overdue data includes: customer information, loan information, current number of overdue days, current repayment status, and sample size; Based on the overdue data, a multi-dimensional dynamic weight is calculated to match the customer's current overdue status; the multi-dimensional dynamic weight includes: the importance weight of overdue days and the sample weight. The multi-dimensional dynamic weights are integrated and normalized to obtain normalized weights; The normalized weights and the overdue data are input into a pre-deployed risk prediction model for inference processing to generate a real-time risk score for the customer, so as to generate a corresponding collection strategy based on the real-time risk score of the customer.

[0007] In one embodiment, the sample size includes: the total number of samples corresponding to the current overdue days among all overdue users and the number of positive samples.

[0008] The sample weights include: class imbalance weights, sample scarcity weights, and label-based compensation factors.

[0009] In one embodiment, the method for calculating the class imbalance weights includes: Calculate the ratio of the number of positive samples to the total number of samples to obtain the proportion of positive samples under the current overdue days; Based on the positive sample ratio, a calculation method that is inversely correlated with the positive sample ratio is adopted, and a preset smoothing term is introduced to avoid calculation abnormalities, thereby obtaining the class imbalance weight.

[0010] In one embodiment, the method for calculating the sample scarcity weight includes: The total sample size corresponding to each overdue day among all overdue users is calculated to obtain the total sample size for each overdue day. Determine the maximum sample size from the total number of samples for each overdue day; Calculate the difference between the total number of samples corresponding to the current number of overdue days and the maximum sample size; Based on the difference, the sample scarcity weight is obtained by using a calculation method that is positively correlated with the difference.

[0011] In one embodiment, the method for calculating the label-based compensation factor includes: Obtain the sample tags corresponding to the customer; the sample tags include: positive sample tags and negative sample tags; the positive sample tags correspond to samples that have been overdue for N days, and the negative sample tags correspond to samples that have not been overdue for N days; N is a preset positive integer; If the sample label is a negative sample label, a fixed compensation factor value is assigned; If the sample label is a positive sample label, obtain the proportion of positive samples corresponding to the current overdue days; A dynamic compensation function is constructed based on the positive sample ratio and the preset smoothing factor. The label-based compensation factor is obtained by combining the results of the dynamic compensation function with the preset compensation intensity coefficient and using an overlay calculation method.

[0012] In one embodiment, the method for calculating the importance weight of the overdue days includes: Based on the current number of overdue days, the importance weight of the overdue days is obtained by using a linearly increasing calculation method.

[0013] In one embodiment, before calculating the multi-dimensional dynamic weights matching the customer's current overdue status based on the overdue data, the method further includes: Determine whether the current number of overdue days belongs to the post-loan M1 overdue stage; If so, perform the step of calculating the multi-dimensional dynamic weight that matches the customer's current overdue status based on the overdue data; If not, end the current process.

[0014] At least one embodiment of this application also provides a post-loan risk dynamic assessment device, characterized in that it includes: The data collection unit is used to collect real-time overdue data of customers; the overdue data includes: customer information, loan information, current number of overdue days, current repayment status, and sample size; The weight calculation unit is used to calculate a multi-dimensional dynamic weight that matches the customer's current overdue status based on the overdue data; the multi-dimensional dynamic weight includes: the importance weight of overdue days and the sample weight; The weight fusion unit is used to fuse the multi-dimensional dynamic weights and perform normalization processing to obtain normalized weights. The risk prediction unit is used to input the normalized weights and the overdue data into a pre-deployed risk prediction model for inference processing to generate a real-time risk score for the customer, so as to generate a corresponding collection strategy based on the real-time risk score for the customer.

[0015] At least one embodiment of this application also provides an electronic device, including: at least one processor; and a memory communicatively connected to said at least one processor; The memory stores instructions that can be executed by the at least one processor, which are then executed by the at least one processor to enable the at least one processor to perform the post-loan risk dynamic assessment method.

[0016] At least one embodiment of this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the post-loan risk dynamic assessment method.

[0017] The post-loan risk dynamic assessment method provided in this application collects customer overdue data in real time and dynamically calculates the weights matching the current overdue status and the corresponding sample weights. This breaks the limitations of traditional static point-in-time scoring and allows for timely tracking of risk changes during the customer's overdue process. The multi-dimensional dynamic weight design fully considers the risk priority of overdue days and the characteristics of sample distribution. It highlights the importance of high-risk customer groups through the weight of overdue days and balances sample differences at different overdue stages through sample weights, effectively solving the long-tail problem of scarce samples from long-term overdue customer groups and avoiding model bias towards certain groups. The model addresses overfitting or underfitting of samples, ensuring that both short-term and long-term delinquency risk groups receive sufficient attention. It then integrates and normalizes multi-dimensional weights to eliminate scale differences and extreme value interference, preserving the relative importance differences between samples. Finally, the real-time risk score generated through model inference serves as the core basis for formulating differentiated collection strategies, enabling precise implementation of high-risk, high-intensity collection and low-risk, gentle reminders. This improves the efficiency of collection resource utilization, reduces ineffective collection costs, optimizes customer experience, and effectively promotes the refined upgrading of post-loan risk management. Attached Figure Description

[0018] One or more embodiments are illustrated by way of example with reference to the accompanying drawings, and these illustrative descriptions do not constitute a limitation on the embodiments.

[0019] Figure 1 This is a flowchart of a post-loan risk dynamic assessment method provided in one embodiment of this application; Figure 2 This is a schematic diagram illustrating a weight calculation process provided in one embodiment of this application; Figure 3 This is a schematic diagram of the module division of a post-loan risk dynamic assessment device provided in one embodiment of this application; Figure 4 This is a schematic diagram of the structure of an electronic device provided in one embodiment of this application. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the various embodiments of this application will be described in detail below with reference to the accompanying drawings. However, those skilled in the art will understand that many technical details have been provided in the various embodiments of this application to help readers better understand this application. However, the technical solutions claimed in this application can be implemented even without these technical details and various changes and modifications based on the following embodiments. The division of the various embodiments below is for the convenience of description and should not constitute any limitation on the specific implementation of this application. The various embodiments can be combined with and referenced by each other without contradiction.

[0021] The following is a detailed description of the implementation details of the post-loan risk dynamic assessment method in this embodiment. The following content is only for the convenience of understanding and is not necessary for implementing this solution.

[0022] Example 1: The specific process of the post-loan risk dynamic assessment method in this embodiment can be as follows: Figure 1 As shown, it includes: Step 101: Collect real-time overdue data from customers.

[0023] We continuously and dynamically collect key information related to customer delinquency, specifically including five core dimensions: customer information, loan information, current number of delinquent days, current repayment status, and sample size.

[0024] The customer information includes core identity and credit-related information such as identity identifiers and basic credit history records; loan information includes, but is not limited to, key data related to the loan contract such as loan amount, disbursement date, repayment plan, and remaining principal; the current overdue days are accurately counted to the time of collection, calculated in calendar days (i.e., the cumulative number of days from the day after the due date to the collection date); the current repayment status is clearly marked as specific categories such as repaid, unpaid, and partially repaid, serving as the core basis for sample labeling; the sample size is categorized and counted according to dimensions such as the same overdue days and the same repayment status, recording the sample size of the corresponding group.

[0025] Customer information and loan information are the foundation for identifying individual customer differences and loan attribute differences. Current overdue days and repayment status are the core indicators for judging risk level. Sample size is a key reference for solving the data sparsity problem. These five types of data together constitute the complete data foundation for risk prediction.

[0026] Step 102: Based on the overdue data, calculate the multi-dimensional dynamic weights that match the customer's current overdue status.

[0027] The overdue data collected in step 101, such as customer information, loan information, current overdue days, current repayment status, and sample size, are linked and integrated. Based on the customer's current overdue status, and in accordance with the preset business rules and weight design ideas, two types of core weights are calculated: the importance weight of overdue days and the sample weight.

[0028] Among them, the importance weight of overdue days is combined with the actual situation of the customer's current overdue days to reflect the risk attention priority corresponding to different overdue durations; the sample weight is combined with the sample distribution characteristics under the current overdue status to realize differentiated consideration of different samples. The two types of weights together constitute a multi-dimensional dynamic weight system.

[0029] It should be noted that this embodiment does not limit the specific calculation algorithm for the importance weight of overdue days and the sample weight. For the importance weight of overdue days, for example, a linear function can be used to quantify the positive correlation between overdue days and risk attention, and the weight value can be increased in an increasing gradient according to the overdue days to reflect the business logic that the longer the overdue period, the higher the risk and the higher the attention priority. Alternatively, a piecewise function can be used to set the weight coefficient for different overdue intervals, and the weight corresponding to key overdue nodes (such as 15 days overdue, 30 days overdue, etc.) can be adjusted in a stepwise manner to adapt to the differences in risk characteristics at different overdue stages.

[0030] Regarding sample weights, for example, inverse proportional weights can be designed based on the proportion of samples in the same overdue status, assigning higher weights to groups with smaller sample sizes to compensate for data scarcity; alternatively, the weight ratio of positive and negative samples in the current overdue status can be adjusted by using a balancing coefficient to avoid the model being overly biased towards one type of sample, thus achieving differentiated consideration of samples with different distribution characteristics. This embodiment only uses the above methods as examples, but is not limited to them.

[0031] The weight calculation is matched in real time with the customer's current delinquency status, enabling the model to promptly capture dynamic changes in customer risk. Simultaneously, it is matched in real time with the sample status, allowing the model to accurately learn risk patterns at different delinquency stages and deeply uncover hidden temporal risk characteristics within the samples. Compared to static weight models, dynamic weights are better adapted to the temporal evolution of post-loan risk, resulting in risk predictions that better align with actual business scenarios. Furthermore, compared to the uniform weighting of traditional models, dynamic weights can be flexibly adjusted based on the customer's current delinquency status, allowing the model to assign appropriate levels of attention to customers with different delinquency durations and sample characteristics. This aligns with the business objective of balancing short-term and long-term risks. The multi-dimensional weight design, including the importance weight of delinquency days and sample weights, fully adapts to the distribution characteristics of delinquency data, effectively addressing the challenges posed by sample differences under different delinquency statuses.

[0032] Step 103: Integrate the multi-dimensional dynamic weights and normalize them to obtain normalized weights.

[0033] By integrating the importance weights of overdue days and sample weights through a pre-defined integration logic, such as multiplying the weights to form a comprehensive weight that can reflect the characteristics of customer overdue risk and sample value, the comprehensive weight is then adjusted to a reasonable range through normalization, such as using the 75th quantile. The core principle of normalization is to preserve the relative differences in weights for different customers and different overdue statuses, while constraining the weight values ​​to a unified and reasonable range to avoid extreme outliers caused by some weight values ​​being too large or too small. This ensures that all weights have a consistent scale standard and can be interpreted equally and accurately by the model, ultimately outputting normalized weights with a unified scale that can be directly used for model training.

[0034] After normalization, the normalized weights can be further validated to confirm that the weight values ​​have no abnormal fluctuations, the relative differences meet expectations, and the overall distribution is uniform and reasonable. This ensures that the output normalized weights can be directly used in subsequent model training without additional adjustments. Of course, this validation step can be omitted; this embodiment does not limit this, and the configuration can be tailored to the specific needs of the application scenario.

[0035] Step 104: Input the normalized weights and overdue data into the pre-deployed risk prediction model for inference processing to generate a real-time risk score for the customer, so as to generate a corresponding collection strategy based on the real-time risk score of the customer.

[0036] Normalized weights and overdue data are used as joint inputs to a pre-trained risk prediction model. The model combines the risk priority and sample value reflected by the normalized weights to accurately interpret the risk signals in the overdue data and generate a real-time risk score for the corresponding customer. The score has a clear risk differentiation (e.g., the higher the score, the higher the overdue risk and the lower the repayment probability), which can intuitively reflect the customer's current risk status.

[0037] After obtaining the real-time risk score, the generated real-time risk score can be further classified according to the business's preset classification standards (such as high, medium and low risk levels), and corresponding collection strategies can be matched for customers with different risk levels, thereby improving the efficiency of collection resource utilization.

[0038] In this embodiment, there are no restrictions on the model type and training method for the risk prediction model. For example, a gradient boosting tree model (such as the XGBoost model) can be used, and the model parameters can be iteratively adjusted through the gradient descent optimization algorithm to learn the differentiated characteristics of long-term and short-term delinquency risks using weighted sample data. Alternatively, a deep learning model (such as a neural network or LSTM) can be used, which captures the temporal correlation and nonlinear risk patterns in delinquency data by constructing a multi-layer network structure. Combined with batch training or online learning, it can be adapted to the real-time updated weighted samples and delinquency data to ensure the model's adaptability to dynamic risks.

[0039] Based on the above introduction, the post-loan risk dynamic assessment method provided in this embodiment, by collecting customer overdue data in real time and dynamically calculating the weights matching the current overdue status and the corresponding sample weights, breaks through the limitations of traditional static point-in-time scoring. It can promptly track risk changes during the customer's overdue process. The multi-dimensional dynamic weight design fully considers the risk priority of overdue days and the sample distribution characteristics. It highlights the attention value of high-risk customer groups through the importance weight of overdue days, and balances the sample differences at different overdue stages through sample weights, effectively solving the long-tail problem of scarce samples for long-term overdue customer groups and avoiding model bias. By addressing overfitting or underfitting in some samples, the model ensures that both short-term and long-term delinquency risk groups receive sufficient attention. Then, the multi-dimensional weights are integrated and normalized to eliminate scale differences and extreme value interference between different dimensions, preserving the relative importance differences between samples. Finally, the real-time risk score generated through model inference can directly serve as the core basis for formulating differentiated collection strategies, enabling precise implementation of high-risk, high-intensity collection and low-risk, gentle reminders. This improves the efficiency of collection resource utilization, reduces ineffective collection costs, optimizes customer experience, and effectively promotes the refined upgrading of post-loan risk management.

[0040] Example 2: In the practice of the above-mentioned post-loan risk dynamic assessment method, considering the significant distribution heterogeneity of overdue data in post-loan scenarios, as the number of overdue days increases, the total sample size gradually decreases, and the proportion of positive samples with repayment status of repayment drops significantly. For example, in some scenarios, the proportion of positive samples in the long overdue stage drops from 60% to less than 2%. The single-dimensional sample weight is difficult to accurately solve the core technical pain points such as category imbalance and sparse long-tail sample data, and it cannot fully guarantee the model's balanced attention to the long-term and short-term overdue risk populations.

[0041] In response to this, this embodiment further refines the definition of sample weights and corresponding sample data. On the one hand, it clarifies that the sample quantity specifically includes the total number of samples and the number of positive samples corresponding to the current overdue days among all overdue users, providing accurate data support for subsequent weight calculation. On the other hand, the sample weights are further refined into category imbalance weights, sample scarcity weights, and label-based compensation factors.

[0042] Specifically, the sample size statistics take the current number of overdue days as the core classification dimension. In the full set of overdue user data, the total number of samples (i.e., the total number of overdue users corresponding to that overdue day) and the number of positive samples (i.e., the number of users who have completed repayment under that overdue day) are counted separately for each overdue day.

[0043] Based on the statistically obtained sample size (total sample size, number of positive samples), the class imbalance weight needs to be quantified by the proportion of positive samples (number of positive samples / total sample size) under different overdue days; the sample scarcity weight needs to be calculated by combining the difference between the total sample size under different overdue days and the largest sample size in the full data; the label-based compensation factor needs to set calculation rules for positive samples (label is 1) and negative samples (label is 0) respectively, focusing on targeted compensation for positive samples. The three types of weights together constitute a complete sample weight system.

[0044] In post-loan scenarios, there are issues such as decreasing total number of samples with different overdue days, a significant disparity in the proportion of positive samples (the proportion of positive samples can drop from 60% to less than 2% in the long-term overdue phase), and sparse long-tail samples. A single-dimensional sample weight cannot comprehensively address these pain points. In this embodiment, sample weights are divided into three categories: class imbalance weights, sample scarcity weights, and label-based compensation factors. Class imbalance weights alleviate the problem of imbalanced positive and negative sample ratios, sample scarcity weights compensate for the sparsity of long-tail sample data, and label-based compensation factors strengthen the focus on key positive samples. The three types of sample weights correspond to different distribution pain points of overdue data, forming a comprehensive weight system of class balance + scarcity compensation + label emphasis. This ensures that the model can accurately identify both long-term and short-term overdue risk groups, avoid overfitting or underlearning of some samples, and improve the balance and accuracy of risk prediction.

[0045] This embodiment does not limit the specific calculation algorithms for class imbalance weights, sample scarcity weights, and label-based compensation factors. For example, an adaptive weighting strategy derived from the SMOTE algorithm can be used to dynamically adjust the class imbalance weights based on the distribution density of positive and negative samples under different overdue days, alleviating class imbalance by increasing the weight ratio of minority class samples. Alternatively, an entropy-based calculation logic can be used, combining the dispersion of sample distribution to quantify sample scarcity, allowing overdue stages with fewer and more dispersed samples to receive higher scarcity weights. For label-based compensation factors, for example, a linear compensation function can be used to set the compensation coefficient based on the repayment probability trend of positive samples under different overdue days, assigning higher compensation weights to positive samples in overdue stages with lower repayment probabilities. Furthermore, a business scenario coefficient can be introduced, and a non-linear compensation model can be designed based on the ratio of collection costs to recovery revenue, ensuring that the model focuses on long-overdue positive samples while also considering actual business revenue objectives.

[0046] Optionally, a method for calculating class imbalance weights includes: (1) Calculate the ratio of the number of positive samples to the total number of samples to obtain the proportion of positive samples under the current overdue days.

[0047] Using the current number of overdue days as a fixed dimension, determine the number of positive samples (i.e., the number of overdue users whose repayment status is "paid") and the total number of samples (i.e., the total number of all overdue users corresponding to the current number of overdue days). Then, by calculating the positive sample ratio under the current number of overdue days, we can obtain the positive sample ratio under the current number of overdue days. This allows us to accurately capture the distribution relationship between positive and negative samples in this overdue period. For example, under a long overdue period, the number of positive samples is small and the positive sample ratio is low, while under a short overdue period, the number of positive samples is large and the positive sample ratio is high.

[0048] (2) Based on the proportion of positive samples, a calculation method that is inversely related to the proportion of positive samples is adopted, and a preset smoothing term is introduced to avoid calculation abnormalities, thus obtaining the class imbalance weight.

[0049] The calculation method that is inversely correlated with the proportion of positive samples means that the lower the proportion of positive samples (i.e., the fewer users who have repaid under the overdue days), the higher the calculated weight, thereby increasing the importance of the minority positive samples in model training and avoiding the model from being overly biased towards the high proportion of negative samples.

[0050] In the calculation method that is inversely correlated with the proportion of positive samples, a preset smoothing term is introduced, that is, a fixed small value (such as 0.01) is added to avoid abnormal situations such as infinite values ​​in the calculation results due to the extremely low proportion of positive samples (approaching 0), thus ensuring the stability and rationality of the weight calculation.

[0051] Ultimately, the category imbalance weights obtained through this logical calculation can effectively balance the problem of the disparity in the proportion of positive and negative samples under different overdue days, allowing the model to learn the risk characteristics of various types of samples fairly.

[0052] Specifically, the formula for calculating the class imbalance weight is as follows: W_imbalance(d)=1 / (p(d)+0.01), where p(d) is the proportion of positive samples corresponding to the overdue days d.

[0053] Optionally, a method for calculating sample scarcity weights includes: (1) Calculate the total number of samples corresponding to each overdue day among all overdue users to obtain the total number of samples for each overdue day.

[0054] Using the number of overdue days as the sole classification dimension, we traverse all overdue user data and count the total number of samples (i.e., the number of all overdue users under that number of overdue days) for each overdue day (e.g., 1 day, 2 days, ... 31 days overdue). This ensures that each overdue day has a clearly corresponding sample size data, with no omissions or duplicates.

[0055] (2) Determine the maximum sample size from the total number of overdue days.

[0056] The total sample size corresponding to each overdue day is compared and ranked to identify the peak data and establish a benchmark value for judging sample scarcity. This benchmark value can quantify the relative scarcity of the total sample size of other overdue days, providing a key reference standard for subsequent weight calculation.

[0057] (3) Calculate the difference between the total number of samples corresponding to the current overdue days and the maximum sample size.

[0058] Using the current number of overdue days as the target dimension, extract the total number of samples (N(d)) corresponding to this dimension, and subtract the total number of samples with the current number of overdue days (N(d)) from the maximum number of samples (N_max) to obtain the difference between the two.

[0059] The scarcity of samples is transformed into a quantifiable numerical indicator. The larger the difference, the greater the gap between the current total number of overdue days and the maximum number of samples, and the scarcer the sample is; the smaller the difference, the more abundant the sample is.

[0060] (4) Based on the difference, the sample scarcity weight is obtained by using a calculation method that is positively correlated with the difference.

[0061] The calculation method, which is positively correlated with the difference, means that the larger the difference (i.e., the scarcer the current overdue days sample), the higher the calculated sample scarcity weight, thereby increasing the importance of scarce samples in model training and avoiding insufficient learning of risk characteristics by the model due to a small sample size. Conversely, the smaller the difference (the more abundant the sample), the lower the weight, preventing abundant samples from excessively dominating model training. Ultimately, the sample scarcity weight obtained through this logic can specifically address the long-tail data sparsity problem where the proportion of long-term overdue customer samples is extremely low (only about 17% of the short-term overdue customer samples), achieving intelligent compensation for scarce samples and ensuring that the model learns the risk characteristics of both long-term and short-term overdue samples in a balanced manner.

[0062] Specifically, a formula for calculating the sample scarcity weight is as follows: W_scarcity(d)=1+(N_max-N(d)) / N_max, where N(d) is the number of samples with overdue days d, and N_max is the maximum sample size.

[0063] Optionally, a method for calculating a label-based compensation factor includes: (1) Obtain the sample labels corresponding to the customers.

[0064] The sample labels include positive sample labels and negative sample labels. Positive sample labels correspond to samples that are overdue for N days and have been repaid, while negative sample labels correspond to samples that are overdue for N days and have not been repaid. That is, if a customer makes a payment within N days of being overdue, it is marked as a positive sample label; if a customer does not make a payment within N days of being overdue, it is marked as a negative sample label. Here, N is a preset fixed positive integer, such as DPD15 commonly used in business scenarios, i.e., N=15.

[0065] (2) If the sample label is a negative sample label, assign a fixed compensation factor value.

[0066] When a sample label is determined to be a negative sample label (i.e., overdue for N days without repayment), it is directly assigned a preset fixed compensation factor value (such as 1.0 by default in the technical solution).

[0067] Negative samples do not require additional compensation. The focus of weight optimization is on positive samples, ensuring that the compensation logic is focused on the core objective of increasing the attention given to long-overdue positive samples, thus avoiding resource waste or imbalance in weight allocation.

[0068] (3) If the sample label is a positive sample label, obtain the proportion of positive samples corresponding to the current overdue days.

[0069] When the sample label is a positive sample label (i.e., repayment within N days of overdue payment), it is necessary to extract the proportion of positive samples under the current overdue days of the sample, that is, the proportion of positive samples (repayment within N days of overdue payment) in the total number of samples corresponding to the current overdue days among all overdue users (p(d)), in order to obtain data reflecting the scarcity of positive samples in the current overdue stage and ensure that the compensation factor matches the distribution characteristics of positive samples.

[0070] (4) Construct a dynamic compensation function based on the proportion of positive samples and a preset smoothing factor.

[0071] Using the positive sample ratio (p(d)) obtained in step (3) as the core variable, a preset smoothing factor ε (usually 0.001) is introduced to construct a dynamic compensation function (e.g., f(y,d)=log(1+1 / (p(d)+ε))).

[0072] The lower the proportion of positive samples (a characteristic of long-term overdue periods), the larger the function result, and the stronger the corresponding compensation demand. At the same time, the smoothing factor can avoid calculation anomalies caused by the proportion of positive samples approaching 0, ensuring the stability and rationality of the function output.

[0073] (5) Combine the results of the dynamic compensation function with the preset compensation intensity coefficient, and use the superposition calculation method to obtain the label-based compensation factor.

[0074] First, determine the preset compensation intensity coefficient λ (the value range can be [0.5, 5.0], which is used to adjust the compensation intensity). Then, multiply the result of the dynamic compensation function by λ, and finally superimpose it with a fixed benchmark value (such as 1.0) to obtain the label-based compensation factor (such as C(y)=1+λ×f(y,d)) corresponding to the positive sample.

[0075] By superimposing calculations to amplify the weight of long-overdue positive samples, the proportion of positive samples in the long-overdue stage is low, the function result is large, and the compensation factor is higher after superimposing λ, thereby significantly increasing the importance of such scarce positive samples in model training, ensuring that the model captures key repayment signals, and avoiding insufficient learning caused by the scarcity of positive samples.

[0076] Specifically, a formula for calculating a label-based compensation factor is as follows:

[0077] Where y is the sample label (1 = repayment, 0 = non-repayment), λ is the compensation intensity coefficient, with a value range of [0.5, 5.0], f(y,d): dynamic compensation function based on the number of overdue days, f(y,d)=log(1+1 / (p(d)+ε)), p(d) is the proportion of positive samples corresponding to the number of overdue days d, and ε is the smoothing factor (which can be 0.001).

[0078] It should be noted that this embodiment only uses the above calculation method as an example for introduction, but it is not limited to this. Other calculation methods can refer to the introduction of this embodiment, and will not be repeated here.

[0079] Example 3: The importance weight of overdue days reflects the risk priority corresponding to different overdue durations. The longer the overdue period, the lower the customer's repayment probability and the higher the risk level, and the higher the attention weight should be assigned by the model. The specific calculation method for the importance weight of overdue days is not limited in the above embodiments. To more accurately and intuitively quantify the positive correlation between overdue duration and risk level, and to ensure that the weight calculation logic is simple and aligns with the core requirements of post-loan risk management, this embodiment proposes a method for calculating the importance weight of overdue days. Specifically, it can be calculated using a linearly increasing method based on the current overdue days to obtain the importance weight of overdue days.

[0080] The core variable in weight calculation is the customer's current number of overdue days. The increase in weight is proportional to the increase in the number of overdue days. From a practical business perspective, an extension of a customer's overdue days usually indicates a continued decline in repayment willingness or a deterioration in repayment ability, with the risk level gradually increasing. Using a linearly increasing weight calculation method allows this increasing risk characteristic to be directly reflected in the weight value. Customers overdue for 1 day have a base weight, and as the number of overdue days increases (e.g., 5 days, 10 days, 30 days), the weight increases proportionally. This ensures that during training and inference, the model prioritizes high-risk customers with long-term overdue periods while maintaining a reasonable weight for low-risk customers with short-term overdue periods. This avoids the problem of high-risk customers being ignored and low-risk customers being over-focused due to a disconnect between weight allocation and risk trends.

[0081] Specifically, the formula for calculating the importance weight of overdue days is as follows: W_importance(d)=1.0+0.3×(d-1), where d is the number of overdue days.

[0082] To enhance understanding, this embodiment introduces a method for calculating multi-dimensional dynamic weights under the configuration of expected day importance weight, category imbalance weight, sample scarcity weight, and label-based compensation factor. (Refer to...) Figure 2 The diagram shows the calculation process of the weights. The calculation formulas for each weight can be found in the above embodiment and will not be repeated here.

[0083] In the comprehensive weight calculation stage, product integration (W_total=W_importance×W_imbalance×W_scarcity×C(y)) is adopted, which not only preserves the independent value of each dimension's weight, but also achieves information complementarity through cross-validation, avoiding the one-sidedness of a single dimension. The normalization process uses the 75th quantile (W_normalized=W_total / percentile(W_total,75)) instead of the traditional maximum value normalization, effectively avoiding the interference of extreme abnormal weights on the model. For example, the abnormal weights of individual extreme overdue samples will not dominate the training process, while preserving the relative differences between samples, ensuring the fairness and stability of model learning.

[0084] Example 4: The M1 delinquency stage after loan disbursement (1-31 days overdue) is a critical risk intervention window in the early stages of customer delinquency. The risk status at this stage directly affects the efficiency of subsequent repayment and collection costs. Long-term delinquent customers beyond the M1 stage (such as those overdue for more than 31 days) have often formed stable delinquency behavior, with solidified risk characteristics and an extremely low probability of repayment.

[0085] To avoid wasting computing resources and making ineffective investments in collection strategy formulation, and to improve resource utilization efficiency and the accuracy of risk assessment in the core stage, the embodiment further proposes to perform the following steps before calculating the multi-dimensional dynamic weights that match the customer's current overdue status based on overdue data in step 102: determine whether the current overdue days belong to the post-loan M1 overdue stage; if yes, proceed to step 102; if no, end the current process.

[0086] Based on the definition of M1 stage (1-31 days overdue), check whether the customer's current overdue days fall within the 1-31 day range: If the current overdue days are within the 1-31 day range, it is determined to belong to M1 stage, indicating that the customer is still in the early stage of overdue payment, the risk status has not yet solidified, and there is a high potential for repayment. Step 102 and subsequent weight calculation, model reasoning, and other processes should be continued to develop a targeted collection strategy for the customer through accurate risk assessment. If the current overdue days exceed 31 days (i.e., exceed M1 stage), it is determined not to belong to this stage, indicating that the customer has formed a stable overdue behavior, the risk characteristics are solidified, and the probability of repayment is extremely low. There is no need to invest computing resources in complex dynamic weight calculation and risk prediction. The current process can be terminated directly to avoid ineffective computing power consumption and waste of collection resources.

[0087] By using phased screening, we can optimize resource allocation by focusing on key areas and discarding inefficient ones. This ensures that customers in the high-value M1 phase receive comprehensive and accurate risk assessment support, while terminating non-core phase processes significantly improves the efficiency and return on investment of overall post-loan risk assessment, making the methodology more aligned with the actual business management needs.

[0088] Of course, this embodiment takes the M1 stage as an example, but other key research objects can also be configured. This embodiment does not limit this, and can be set accordingly according to actual needs.

[0089] Example 5: This embodiment relates to a post-loan risk dynamic assessment device. A schematic diagram of the post-loan risk dynamic assessment device in this embodiment is shown below. Figure 3 As shown, it includes: a data collection unit 201, a weight calculation unit 202, a weight fusion unit 203, and a risk prediction unit 204.

[0090] The data collection unit 201 is used to collect real-time overdue data from customers. The overdue data includes: customer information, loan information, current number of overdue days, current repayment status, and sample size. The weight calculation unit 202 is used to calculate a multi-dimensional dynamic weight that matches the customer's current overdue status based on overdue data; the multi-dimensional dynamic weight includes: the importance weight of overdue days and the sample weight; The weight fusion unit 203 is used to fuse multi-dimensional dynamic weights and perform normalization processing to obtain normalized weights. The risk prediction unit 204 is used to input normalized weights and overdue data into a pre-deployed risk prediction model for inference processing to generate a real-time risk score for the customer, so as to generate a corresponding collection strategy based on the real-time risk score of the customer.

[0091] In the post-loan risk dynamic assessment device provided in this embodiment, the data collection unit focuses on real-time overdue data, breaking through the limitations of traditional static data collection, and ensuring that key information such as the customer's current overdue days and repayment status are synchronized with risk evolution. The weight calculation unit generates multi-dimensional dynamic weights based on real-time data. Among them, the weight of the importance of overdue days reflects the business logic that the longer the overdue period, the higher the risk. The sample weight is specifically balanced to balance the differences in sample distribution. The combination of the two allows the weight allocation to be deeply matched with the customer's current overdue status, realizing the accurate capture of the customer's latest repayment intention and risk characteristics. The weight fusion unit integrates multi-dimensional dynamic weights through comprehensive calculation and uses normalization processing to eliminate extreme value interference and unify the weight scale. It retains the relative differences of each dimension weight to convey differentiated risk information, and ensures that the output normalized weight can directly adapt to the input requirements of the risk prediction model. The risk prediction unit inputs the normalized weights and overdue data into the pre-deployed model, and the generated real-time risk score can be directly used as a quantitative basis for the formulation of collection strategies.

[0092] The real-time risk score of this device is more business-oriented, enabling financial institutions to develop strong collection plans for high-risk customers and adopt gentle reminder strategies for low-risk customers. It improves the efficiency of collection resource utilization while reducing the cost of ineffective collection, effectively transforming technological advantages into business value and effectively solving the problem of the disconnect between traditional post-loan assessment devices and actual collection scenarios.

[0093] Additionally, it should be noted that all modules involved in this embodiment are logical modules. In practical applications, a logical unit can be a physical unit, a part of a physical unit, or a combination of multiple physical units.

[0094] Furthermore, in order to highlight the innovative aspects of this application, no units that are not closely related to solving the technical problems proposed in this application are introduced in this embodiment, but this does not mean that there are no other units in this embodiment.

[0095] Example 6: Another embodiment of this application relates to an electronic device, such as... Figure 4 As shown, it includes: at least one processor 301; and a memory 302 communicatively connected to at least one processor 301; wherein the memory 302 stores instructions executable by at least one processor 301, the instructions being executed by at least one processor 301 to enable at least one processor 301 to perform the steps of the post-loan risk dynamic assessment method in the above embodiments.

[0096] The memory and processor are connected via a bus, which can include any number of interconnecting buses and bridges, connecting various circuits of one or more processors and memories. The bus can also connect various other circuits, such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and will not be described further herein. The bus interface provides an interface between the bus and the transceiver. The transceiver can be a single element or multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices over a transmission medium. Data processed by the processor is transmitted over the wireless medium via an antenna, which further receives data and transmits it to the processor.

[0097] The processor manages the bus and general processing, and also provides various functions, including timing, peripheral interfaces, voltage regulation, power management, and other control functions. Memory is used to store data used by the processor during operation.

[0098] Example 7: Another embodiment of this application relates to a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the method embodiments described above.

[0099] That is, those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a device (which may be a microcontroller, chip, etc.) or processor to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0100] Those skilled in the art will understand that the above embodiments are specific embodiments for implementing this application, and in practical applications, various changes can be made to them in form and detail without departing from the spirit and scope of this application.

Claims

1. A method for dynamic evaluation of post-loan risks, characterized in that, include: Collect real-time overdue data from customers; The overdue data includes: customer information, loan information, current number of overdue days, current repayment status, and sample size; Based on the overdue data, a multi-dimensional dynamic weight is calculated to match the customer's current overdue status; the multi-dimensional dynamic weight includes: the importance weight of overdue days and the sample weight. The multi-dimensional dynamic weights are integrated and normalized to obtain normalized weights; The normalized weights and the overdue data are input into a pre-deployed risk prediction model for inference processing to generate a real-time risk score for the customer, so as to generate a corresponding collection strategy based on the real-time risk score of the customer.

2. The method for dynamic assessment of post-loan risks according to claim 1, characterized in that, The sample size includes: the total number of samples corresponding to the current overdue days among all overdue users and the number of positive samples. The sample weights include: class imbalance weights, sample scarcity weights, and label-based compensation factors.

3. The method for dynamic assessment of post-loan risks according to claim 2, characterized in that, The method for calculating the class imbalance weights includes: Calculate the ratio of the number of positive samples to the total number of samples to obtain the proportion of positive samples under the current overdue days; Based on the positive sample ratio, a calculation method that is inversely correlated with the positive sample ratio is adopted, and a preset smoothing term is introduced to avoid calculation abnormalities, thereby obtaining the class imbalance weight.

4. The method for dynamically assessing credit risk according to claim 2, wherein, The method for calculating the sample scarcity weight includes: The total sample size corresponding to each overdue day among all overdue users is calculated to obtain the total sample size for each overdue day. Determine the maximum sample size from the total number of samples for each overdue day; Calculate the difference between the total number of samples corresponding to the current number of overdue days and the maximum sample size; Based on the difference, the sample scarcity weight is obtained by using a calculation method that is positively correlated with the difference.

5. The post-loan risk dynamic assessment method according to claim 2, characterized in that, The method for calculating the label-based compensation factor includes: Obtain the sample tags corresponding to the customer; the sample tags include: positive sample tags and negative sample tags; the positive sample tags correspond to samples that have been overdue for N days, and the negative sample tags correspond to samples that have not been overdue for N days; N is a preset positive integer; If the sample label is a negative sample label, a fixed compensation factor value is assigned; If the sample label is a positive sample label, obtain the proportion of positive samples corresponding to the current overdue days; A dynamic compensation function is constructed based on the positive sample ratio and the preset smoothing factor. The label-based compensation factor is obtained by combining the results of the dynamic compensation function with the preset compensation intensity coefficient and using an overlay calculation method.

6. The post-loan risk dynamic assessment method according to claim 1, characterized in that, The calculation method for the importance weight of the overdue days includes: Based on the current number of overdue days, the importance weight of the overdue days is obtained by using a linearly increasing calculation method.

7. The post-loan risk dynamic assessment method according to claim 1, characterized in that, Before calculating the multi-dimensional dynamic weights that match the customer's current overdue status based on the overdue data, the method further includes: Determine whether the current number of overdue days belongs to the post-loan M1 overdue stage; If so, perform the step of calculating the multi-dimensional dynamic weight that matches the customer's current overdue status based on the overdue data; If not, end the current process.

8. A post-loan risk dynamic assessment device, characterized in that, include: The data collection unit is used to collect customers' real-time overdue data; The overdue data includes: customer information, loan information, current number of overdue days, current repayment status, and sample size; The weight calculation unit is used to calculate a multi-dimensional dynamic weight that matches the customer's current overdue status based on the overdue data; the multi-dimensional dynamic weight includes: the importance weight of overdue days and the sample weight; The weight fusion unit is used to fuse the multi-dimensional dynamic weights and perform normalization processing to obtain normalized weights. The risk prediction unit is used to input the normalized weights and the overdue data into a pre-deployed risk prediction model for inference processing to generate a real-time risk score for the customer, so as to generate a corresponding collection strategy based on the real-time risk score for the customer.

9. An electronic device, characterized in that, include: At least one processor; And, a memory communicatively connected to the at least one processor; The memory stores instructions that can be executed by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform the post-loan risk dynamic assessment method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the post-loan risk dynamic assessment method as described in any one of claims 1 to 7.

Images