An insurance customer refund risk prediction and retention intervention intelligent system

Abstract

The present application relates to a kind of insurance customer's risk prediction and retention intervention intelligent system of withdrawing from insurance, specifically relates to the field of insurance information technology processing, the scheme is integrated by multi-source data fusion module policy information, customer behavior data and external data, overcome the defect that data dimension is limited;Withdrawal risk prediction module adopts gradient boosting decision tree to train withdrawal risk prediction model, according to feature contribution and market fluctuation factor dynamic adjustment grading threshold, realize fine-grained accurate classification;Retention intervention module combines customer historical preference and real-time behavior, through strategy engine matches personalized retention intervention strategy, break through the limitation of single intervention strategy;Closed-loop optimization module is updated according to retention conversion index withdrawal risk prediction model and strategy library, form from prediction to intervention again to feedback closed-loop optimization mechanism, significantly improve withdrawal early warning accuracy and retention conversion success rate.

Description

Technical Field

[0001] This invention relates to the field of insurance information technology processing, and more specifically, to an intelligent system for predicting and retaining insurance customer surrender risks. Background Technology

[0002] In the operation of long-term life insurance products such as long-term life insurance, critical illness insurance, and annuity insurance, insurance companies face increasingly complex scenarios in managing customer surrender risks. Internally, policy information systems, customer service platforms, and payment record systems operate independently, with varying data formats and a lack of real-time integration mechanisms. Policy information, such as premium amount, coverage period, and payment method, as well as customer behavior data, such as policy inquiry frequency, customer service consultation content, and payment timeliness, are scattered across multiple business systems. Externally, competitors frequently launch new products with better cost-effectiveness or greater flexibility, and macroeconomic fluctuations, such as interest rate cuts and rising unemployment, directly affect customers' willingness and ability to pay premiums. Customer surrender behavior is driven by these multiple dynamic factors, and traditional static early warning methods cannot capture risk signals in a timely manner, leading insurance companies to often only respond passively when customers are about to surrender, missing the best opportunity for intervention.

[0003] Currently available early warning solutions for policy cancellations in the industry mainly include statistical regression models based on historical cancellation samples, cluster analysis based on customer profiles, and simple threshold warnings based on rule engines, such as marking two consecutive months of non-payment as high risk. These solutions generally suffer from three technical shortcomings: First, the data dimensions are severely limited, failing to incorporate information about industry competitors, such as their product rates, coverage, and promotional activities, as well as macroeconomic data such as GDP growth, disposable income, and interest rates into the feature space, thus failing to effectively explain cancellation behavior caused by changes in the external environment; Second, the risk classification granularity is too coarse, typically only divided into high, medium, and low levels with fixed classification standards, lacking... Third, intervention strategies lack data-driven personalized matching. Retention interventions mainly rely on the experience and judgment of human customer service, resulting in a single strategy, such as uniformly sending coupons. This fails to differentiate the push based on the actual needs of customers with different risk levels, such as premium pressure, insufficient coverage, or service experience. Furthermore, intervention actions are disconnected from risk prediction results, failing to form a closed-loop optimization mechanism from prediction to intervention to feedback. These technical bottlenecks lead to low accuracy rates for policy cancellation warnings and low success rates for retention conversion. There is an urgent need for an intelligent system that can automatically integrate multi-source heterogeneous data, utilize advanced machine learning algorithms to achieve fine-grained dynamic classification, and automatically push personalized intervention strategies. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing an intelligent system for predicting and intervening in the risk of insurance policy surrender. The system solves the problems mentioned in the background by using a multi-source data fusion module, a surrender risk prediction module, a retention intervention module, and a closed-loop optimization module.

[0005] The technical solution of this invention to solve the above-mentioned technical problems is as follows: specifically including: Multi-source data fusion module: Integrates insurance customer policy information, customer behavior data, and external data. Customer behavior data includes policy inquiry frequency, customer service consultation content, and payment timeliness. External data includes industry competitor information and macroeconomic data. The module performs cleaning, alignment, and normalization through an extraction, transformation, and loading pipeline. It uses natural language processing to extract keywords indicating policy cancellation intentions from customer service consultation content, transforms competitor information into product attractiveness comparison indicators, and aligns macroeconomic data by timestamp to construct a wide-table dataset containing static attributes and dynamic time-series features. The policy surrender risk prediction module uses a gradient boosting decision tree to train a policy surrender risk prediction model based on a wide table dataset, outputting the probability of each customer surrendering their policy within a future time window. It calculates the feature contribution degree based on the contribution of each feature in the wide table dataset to the model output, and calculates the market volatility factor based on macroeconomic data. Based on the feature contribution degree and the market volatility factor, it dynamically adjusts the risk level classification threshold to accurately classify customers and generate a customer list with risk level labels and the probability of policy surrender. Retention Intervention Module: Based on the risk level and customer surrender probability in the customer list, combined with the customer's historical preferences and real-time behavior extracted from the wide table dataset, the module matches personalized retention intervention strategies through the strategy engine, pushes intervention instructions through the multi-channel distribution engine, and records intervention logs. Closed-loop optimization module: It performs correlation analysis with intervention logs, customer surrender probability, product attractiveness comparison indicators, and aligned macroeconomic data. It determines retention conversion indicators based on whether the customer cancels the surrender application, pays on time, or increases the policy limit after intervention. It updates the surrender risk prediction model and the strategy library in the strategy engine based on the retention conversion indicators to achieve closed-loop optimization.

[0006] In a preferred embodiment, the feature construction of customer behavior data in the multi-source data fusion module specifically includes: To determine the frequency of policy inquiries, an exponentially weighted moving average method was used to calculate the short-term and long-term averages, respectively. The difference between the two was used as an indicator of the change in inquiry activity. A positive value of the indicator indicates that the inquiry frequency has increased recently, while a negative value indicates that it has decreased. For customer service inquiries, a pre-trained language model is used to encode them into semantic vectors. The cosine similarity between the vectors and the semantic center vectors of a pre-built keyword library for policy cancellation is calculated. The vectors are then mapped to a cancellation intention strength index between zero and one through a Logistic transformation. The closer the cancellation intention strength index is to one, the stronger the intention to cancel the policy. To address the timeliness of payments, the non-negative portion of the actual number of days of payment delay is nonlinearly compressed using a hyperbolic tangent function to obtain a delay severity index with a value ranging from zero to one. The delay severity index tends to one as the number of days of delay increases.

[0007] In a preferred embodiment, the feature construction of external data in the multi-source data fusion module specifically includes: Based on information about industry competitors, we compare our own products with all competitors one by one, calculate the logarithm of the premium ratio, the difference in coverage, and the weighted Euclidean distance of the promotional efforts, take the minimum value, and then normalize it using the Logistic function to obtain a product attractiveness comparison index with a value between zero and one. The larger the product attractiveness comparison index value, the more obvious the overall attractiveness disadvantage of our own products relative to the competitors corresponding to the minimum value. For the aligned macroeconomic data, after normalizing the range of GDP growth rate, disposable income growth rate and benchmark interest rate, the macroeconomic pressure index is obtained by combining the elasticity coefficients and substituting them into the Logistic function. The macroeconomic pressure index is between zero and one. The closer the macroeconomic pressure index is to one, the greater the macroeconomic pressure.

[0008] In a preferred embodiment, the surrender risk prediction module uses a surrender risk prediction model trained on a gradient boosting decision tree based on a wide table dataset, which is used to output the customer surrender probability within a future time window. The calculation process of feature contribution is as follows: For each feature in the wide table dataset, multiple random perturbations are performed within the value range of each feature in the wide table dataset. During each perturbation, the feature value is replaced with a value randomly sampled from the feature's empirical distribution, while keeping other features unchanged. The feature is then re-input into the surrender risk prediction model to calculate the surrender probability and obtain the perturbed probability sequence. The feature contribution is equal to the average of the absolute values ​​of the difference between the probability of policy cancellation before and after the disturbance. The larger the average value, the more significant the influence of the feature on the probability of policy cancellation. The top five features in terms of contribution and their contribution values ​​are selected for dynamic adjustment of subsequent classification thresholds.

[0009] In a preferred embodiment, the calculation process of the market volatility factor in the surrender risk prediction module is as follows: Based on the time series of the macroeconomic pressure index output by the multi-source data fusion module, the macroeconomic pressure index values ​​of the most recent four quarters are taken, the first difference between adjacent quarters is calculated, and the first difference is then subtracted again to obtain the second difference. The absolute value of the second difference is taken and the weighted sum is calculated. The result is the market volatility factor. A higher market volatility factor value indicates that macroeconomic pressures have fluctuated sharply in the near term, and market uncertainty has increased.

[0010] In a preferred embodiment, the dynamic adjustment process of the tiered threshold in the policy surrender risk prediction module is as follows: Based on the average contribution of the top five features and the market volatility factor, the three basic classification thresholds of extremely high risk, high risk and medium risk are modified; whereby the average contribution of the top five features is equal to the sum of the contributions of the top five features divided by five. The correction method is as follows: subtract the product of the preset market volatility adjustment coefficient and the market volatility factor mapped by the hyperbolic tangent function from each basic grading threshold, and then subtract the product of the preset feature contribution adjustment coefficient and the mean feature contribution to obtain the adjusted grading threshold. The adjusted lower limit of the classification threshold shall not be lower than 80% of the basic classification threshold, so as to achieve dynamic and accurate classification of customer risk levels.

[0011] In a preferred embodiment, the process of the strategy engine matching personalized retention intervention strategies in the retention intervention module is as follows: For each customer, the strategy engine iterates through all its built-in strategy templates and calculates the matching score of each strategy template for the customer. The matching score is obtained by weighted summation of three sub-items: risk suitability, historical preference similarity, and real-time urgency. Risk fit measures the degree of match between the strategy and the client's risk level. It is calculated by mapping the risk level to a numerical value and then calculating the absolute difference between the value and the optimal applicable risk level preset in the strategy template. Historical preference similarity measures the cosine similarity between the customer's historical preferences, represented as a historical preference vector, and the policy feature vector corresponding to the policy template. Real-time urgency is calculated based on the customer's probability of canceling the policy and by representing real-time behavior as a real-time behavior vector through a nonlinear combination function. The strategy with the highest matching score is selected as the optimal intervention strategy for the client.

[0012] In a preferred embodiment, the process of pushing intervention instructions via the multi-channel distribution engine in the retention intervention module is as follows: After selecting the optimal intervention strategy, the strategy engine instantiates specific intervention content based on the parameter rules of the strategy template and the customer's specific information, including the reduction ratio, extended protection items, or exclusive service content. The instantiated intervention instructions are simultaneously pushed to the SMS gateway, APP push service and customer service outbound call system through a multi-channel distribution engine; The selection of push channels is based on the online status indicator in real-time behavior. If the customer is currently online, push will be sent via the APP first. If the customer is offline, they will be reached through SMS and customer service outbound calls in parallel. All push operations are completed within a preset time and an intervention log is recorded. The intervention log includes customer identifier, intervention timestamp, push channel, strategy type, strategy instantiation parameters, customer surrender probability, risk level, matching score, and customer's subsequent feedback behavior.

[0013] In a preferred embodiment, the closed-loop optimization module uses customer identifiers and timestamps as keys to associate intervention logs with customer policy cancellation probability, product attractiveness comparison indicators, and aligned macroeconomic data to construct an extended dataset containing pre-intervention characteristics and post-intervention behavioral results. The process of determining the retention conversion index is as follows: assign different benchmark weights and elasticity coefficients to the three behavioral outcomes after customer intervention, namely whether the cancellation application is withdrawn, whether the premium is paid on time, and whether the policy limit is increased. Calculate the contribution value of each behavioral outcome through a nonlinear exponential function, and then sum the three contribution values ​​to obtain the retention conversion index. Among them, the benchmark weight for withdrawing the policy cancellation application is the highest, followed by timely payment, and the weight for increasing the policy amount is the lowest. The elasticity coefficient controls the saturation rate of the contribution of each behavior to the indicator, so that the marginal contribution of a single behavior to the indicator decreases as the number of times the behavior occurs increases.

[0014] In a preferred embodiment, the process of updating the surrender risk prediction model in the closed-loop optimization module is as follows: Extract the diagonal elements of the Fisher information matrix for each parameter from the current surrender risk prediction model. The diagonal elements measure the importance of each parameter to historical data. Then, a loss function containing the new intervention data is constructed. The loss function consists of the log loss of the new intervention data and the elasticity penalty term based on the diagonal elements of the Fisher information matrix. The elasticity penalty term is the sum of the diagonal elements of the Fisher information matrix multiplied by the square of the difference between the old and new parameters and then multiplied by the regularization strength. By optimizing the loss function, the parameters of the surrender risk prediction model are updated incrementally, absorbing the knowledge contained in the new intervention data while retaining the knowledge learned from the historical data. The process of updating the strategy library in the strategy engine is as follows: For each strategy template, it is first determined whether the customer has withdrawn the cancellation application. If the cancellation application is withdrawn, the intervention is directly determined to be successful. If the cancellation application is not withdrawn, the intervention is determined to be successful based on whether the retention conversion index exceeds the preset threshold. The posterior success rate distribution of the strategy template is updated based on the success or failure of the intervention, and then the weight of the strategy template in subsequent matching is adjusted to achieve adaptive evolution of the strategy library.

[0015] The beneficial effects of this invention are as follows: The multi-source data fusion module integrates policy information, customer behavior data, and external data, overcoming the limitation of data dimensions; the surrender risk prediction module uses a gradient boosting decision tree to train the surrender risk prediction model, dynamically adjusting the grading threshold based on feature contribution and market volatility factors to achieve fine-grained and accurate grading; the retention intervention module combines customer historical preferences and real-time behavior, matching personalized retention intervention strategies through a strategy engine, breaking through the limitation of a single intervention strategy; and the closed-loop optimization module updates the surrender risk prediction model and strategy library based on retention conversion indicators, forming a closed-loop optimization mechanism from prediction to intervention to feedback, significantly improving the accuracy of surrender warnings and the success rate of retention conversion. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a block diagram of the system structure of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0019] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0020] Example 1: This example provides the following... Figure 1-2 The system shown is an intelligent system for predicting and intervening in the risk of insurance policy surrender, specifically including: Multi-source data fusion module: Integrates insurance customer policy information, customer behavior data, and external data. Customer behavior data includes policy inquiry frequency, customer service consultation content, and payment timeliness. External data includes industry competitor information and macroeconomic data. The module performs cleaning, alignment, and normalization through an extraction, transformation, and loading pipeline. It uses natural language processing to extract keywords indicating policy cancellation intentions from customer service consultation content, transforms competitor information into product attractiveness comparison indicators, and aligns macroeconomic data by timestamp to construct a wide-table dataset containing static attributes and dynamic time-series features. The policy surrender risk prediction module uses a gradient boosting decision tree to train a policy surrender risk prediction model based on a wide table dataset, outputting the probability of each customer surrendering their policy within a future time window. It calculates the feature contribution degree based on the contribution of each feature in the wide table dataset to the model output, and calculates the market volatility factor based on macroeconomic data. Based on the feature contribution degree and the market volatility factor, it dynamically adjusts the risk level classification threshold to accurately classify customers and generate a customer list with risk level labels and the probability of policy surrender. Retention Intervention Module: Based on the risk level and customer surrender probability in the customer list, combined with the customer's historical preferences and real-time behavior extracted from the wide table dataset, the module matches personalized retention intervention strategies through the strategy engine, pushes intervention instructions through the multi-channel distribution engine, and records intervention logs. Closed-loop optimization module: It performs correlation analysis with intervention logs, customer surrender probability, product attractiveness comparison indicators, and aligned macroeconomic data. It determines retention conversion indicators based on whether the customer cancels the surrender application, pays on time, or increases the policy limit after intervention. It updates the surrender risk prediction model and the strategy library in the strategy engine based on the retention conversion indicators to achieve closed-loop optimization.

[0021] In this embodiment, it is specifically necessary to explain that the feature construction of customer behavior data in the multi-source data fusion module includes: To determine the frequency of policy inquiries, an exponentially weighted moving average method is used to calculate both the short-term and long-term mean values. The difference between these two values ​​is used as an indicator of change in inquiry activity. A positive value indicates a recent increase in inquiry frequency, while a negative value indicates a decrease. The calculation process for the exponentially weighted moving average method is as follows: For each customer, the daily policy inquiry frequency values ​​are processed sequentially according to time. A smoothing coefficient of 0.3 is used for the short-term mean, and a smoothing coefficient of 0.05 is used for the long-term mean. Specifically, the short-term mean at the first time point is initialized to equal the inquiry frequency value at that time point, and then starting from the second time point... The short-term mean at each point in time is equal to 0.3 times the current query frequency plus 0.7 times the short-term mean of the previous point in time. The long-term mean is calculated in the same way, except that the smoothing coefficient is replaced with 0.05, that is, the long-term mean at each point in time is equal to 0.05 times the current query frequency plus 0.95 times the long-term mean of the previous point in time. Finally, the query activity change index is equal to the short-term mean minus the long-term mean, with a value between -10 and 10. An index greater than zero indicates that customers' recent query behavior is more frequent than before, which may indicate an increased intention to cancel the policy; an index less than zero indicates that query behavior is becoming more level. For customer service inquiries, a pre-trained language model encodes the inquiries into semantic vectors. The cosine similarity is then calculated between these vectors and the semantic center vectors of a pre-built keyword lexicon for policy cancellation. This cosine similarity is then mapped using a Logistic transformation to a cancellation intent strength index between zero and one. An index closer to one indicates a stronger cancellation intent. The pre-trained language model uses a BERT model based on the Transformer architecture. This model is pre-trained on a large-scale Chinese corpus for masked language modeling and next-sentence prediction tasks. The training corpus contains approximately three billion Chinese words, covering various text styles including news, encyclopedias, and social media. During use, the text content of each customer inquiry is input into the BERT model. The output vector corresponding to the [CLS] marker in the last hidden layer is taken as the semantic vector of the text, with a vector dimension of 768. The policy cancellation keyword lexicon consists of one hundred seed words highly relevant to policy cancellation behavior, including "cancellation," "cooling-off period," "cash value," "loss," and "contract termination." The semantic center vectors of these words are obtained by averaging the BERT encodings of each seed word. The calculation process of cosine similarity is as follows: perform a dot product operation between the semantic vector and the semantic center vector, and then divide by the product of their magnitudes to obtain a similarity value between -1 and 1; the formula of Logistic transformation is: the withdrawal intention intensity index is equal to 1 divided by 1 plus e to the power of -5 multiplied by the cosine similarity minus 0.3; the index is exactly 0.5 when the cosine similarity is 0.3, the index approaches 1 when the cosine similarity approaches 1, and the index drops rapidly to close to zero when the cosine similarity is less than 0.3; Regarding payment timeliness, the non-negative portion of the actual payment delay days is non-linearly compressed using a hyperbolic tangent function to obtain a delay severity index ranging from zero to one. The delay severity index approaches one as the delay days increase. The actual payment delay days refer to the difference between the customer's actual payment date and the agreed payment date, in days. If the customer pays in advance, the value is negative. The non-negative portion is obtained as follows: if the delay days are less than zero, it is taken as zero; otherwise, the delay days themselves are taken. The mathematical definition of the hyperbolic tangent function is: the hyperbolic tangent value is equal to the difference between e raised to the power of x and e raised to the power of -x, divided by the sum of e raised to the power of x and e raised to the power of -x, where x is equal to the non-negative delay days divided by thirty. This function maps the input value to an open interval between zero and one. When the non-negative delay days are zero, the output is zero; when the non-negative delay days are thirty, the output is approximately 0.76; and when the non-negative delay days reach sixty, the output approaches 0.99. This indicator reflects the severity of payment delays; the higher the value, the worse the customer's payment habits, and the higher the risk of policy cancellation. In the multi-source data fusion module, the feature construction of external data specifically includes: For competitor information, we compare our own products with all competitors one by one, calculating the weighted Euclidean distance of the logarithm of the premium ratio, the difference in coverage, and the promotional intensity. The minimum value is then normalized using a Logistic function to obtain a product attractiveness comparison index ranging from zero to one. A higher product attractiveness comparison index indicates a more significant disadvantage in overall attractiveness for our own products compared to the competitor with the minimum value. The method for obtaining competitor information is as follows: We use a web crawler to regularly access mainstream insurance product comparison websites, the State Administration of Financial Supervision and Management's filing database, and the official websites of major insurance companies on a weekly basis to capture publicly available information on long-term life insurance, critical illness insurance, and annuity insurance products of the same type as our own products, including premium amounts, coverage descriptions, and promotional details. The unstructured text captured is then processed using regular expressions and named entity recognition technology to extract structured fields. The calculation process for the logarithm of the premium ratio is as follows: take the natural logarithm of the premium of the proprietary product and subtract the natural logarithm of the premium of the competitor, then square it; the calculation process for the difference in coverage is as follows: subtract the coverage rate of the competitor from the coverage rate of the proprietary product, then square it; the promotional intensity is directly taken as the square of the discount ratio of the competitor; the three squared terms are multiplied by their corresponding weights, and then added together, with weights of 0.5, 0.3, and 0.2 respectively, and then the square root is taken to obtain the weighted Euclidean distance; after traversing all competitors, the minimum weighted Euclidean distance value is taken, the offset constant 0.2 is subtracted, and then substituted into the Logistic function: the product attractiveness comparison index is equal to one divided by one plus e to the power of negative one multiplied by the difference; this index is exactly 0.5 when the minimum weighted Euclidean distance is equal to 0.2, and when the minimum weighted Euclidean distance is much greater than 0.2, the index approaches one, indicating that the proprietary product is significantly lagging behind a certain competitor in terms of premium, coverage, or promotional intensity, and customers may therefore switch to the competitor; For the aligned macroeconomic data, after normalizing the range of GDP growth rate, disposable income growth rate, and benchmark interest rate, a macroeconomic pressure index with a value between zero and one is obtained by combining elasticity coefficients and substituting them into the Logistic function. The closer the macroeconomic pressure index is to one, the greater the macroeconomic pressure. The macroeconomic data is obtained by subscribing to the quarterly GDP growth rate and disposable income growth rate data released by the National Bureau of Statistics through an API interface, as well as the loan prime rate released by the People's Bank of China as the benchmark interest rate. The data is updated quarterly and aligned with the observation time point of the customer's insurance policy according to the timestamp. The range normalization calculation process is as follows: For each indicator, take the maximum and minimum values ​​of that indicator in all historical data, subtract the minimum value from the current value, and divide by the difference between the maximum and minimum values, so that the result falls between zero and one; the elasticity coefficients are as follows: the elasticity coefficient corresponding to GDP growth rate is +2.0, the elasticity coefficient corresponding to the growth rate of disposable income is +1.5, and the elasticity coefficient corresponding to the benchmark interest rate is +1.0; the elasticity coefficients are a set of preferred values, which can be adjusted with positive weights by those skilled in the art, as long as the combined value increases monotonically as the positive indicator decreases and the negative indicator increases; The calculation process for the combined value is as follows: Subtract the normalized GDP growth rate from 1, multiply by 2.0, add the difference between 1 and the normalized disposable income growth rate, multiply by 1.5, and add the normalized benchmark interest rate, multiplied by 1.0, to obtain the combined value. The combined value ranges from a minimum of 0 to a maximum of 4.5. Finally, substitute the combined value into the Logistic function: the macroeconomic stress index equals 1 divided by 1 plus e raised to the power of the negative combined value. When the combined value reaches its maximum of 4.5, the macroeconomic stress index is approximately 0.989, approaching 1; when the combined value is zero, the macroeconomic stress index is 0.5. Since GDP growth rate and disposable income growth rate are positive indicators, their decline will increase the difference between 1 and the normalized value, thus increasing the combined value and raising the stress index. The benchmark interest rate is a negative indicator; its increase will increase the normalized value, similarly increasing the combined value and raising the stress index. The higher the index, the more unfavorable the macroeconomic environment is for customers to maintain their policies, and the higher the risk of policy surrender.

[0022] In this embodiment, it is specifically noted that in the policy cancellation risk prediction module, the policy cancellation risk prediction model is trained using a gradient boosting decision tree based on a wide table dataset. This model is used to output the probability of each customer canceling their policy within a future time window. The training process of the gradient boosting decision tree is as follows: First, the wide table dataset is divided into a training set and a validation set in an 8:2 ratio. The model parameters are set as follows: the learning rate is 0.1, the maximum tree depth is six layers, the subsampling ratio is 0.8, the feature sampling ratio is 0.7, the minimum number of samples per leaf node is twenty, the maximum number of iterations is five hundred, and an early stopping method is used, stopping training when the validation set loss no longer decreases for ten consecutive iterations. During training, each iteration fits a new decision tree based on the residual of the previous iteration, and the predicted values ​​of all trees are summed to obtain the final policy cancellation probability output. The future time window is set to ninety days, meaning the model outputs the probability value of a customer canceling their policy within the next ninety days, with a value ranging from zero to one. The calculation process for feature contribution is as follows: For each feature in the wide table dataset, multiple random perturbations are performed within the value range of each feature in the wide table dataset. During each perturbation, the feature value is replaced with a value randomly sampled from the feature's empirical distribution, while keeping other features unchanged. The result is then re-input into the policy surrender risk prediction model to calculate the surrender probability, yielding a perturbed probability sequence. The specific implementation of the random perturbation is as follows: For each feature, its empirical distribution in the wide table dataset is first calculated, i.e., the frequency histogram of all values ​​of the feature. Then, the feature is subjected to one hundred random perturbations. During each perturbation, a value is randomly sampled from the empirical distribution and used to replace the original value of the current sample for that feature, while keeping other features unchanged. The perturbed feature vector is then input into the policy surrender risk prediction model to obtain a new surrender probability value. This process is repeated one hundred times to obtain one hundred perturbed probability values. The number of perturbations is set to one hundred to ensure statistical stability. The feature contribution is equal to the average of the absolute values ​​of the difference between the policy cancellation probabilities before and after the perturbation. The larger the average value, the more significant the influence of the feature on the customer's policy cancellation probability. The calculation process for the absolute value of the difference between the policy cancellation probabilities before and after the perturbation is as follows: For each perturbation, calculate the absolute value of the difference between the original policy cancellation probability and the policy cancellation probability after the perturbation, then sum the absolute values ​​one hundred times, and divide by one hundred to obtain the feature contribution for the customer. The contribution value ranges from zero to one, where zero indicates that the feature has no influence on the policy cancellation probability, and one indicates that it completely determines the policy cancellation probability. This indicator reflects the importance of the feature to the model's decision. The top five features by contribution and their contribution values ​​are selected for dynamic adjustment of subsequent grading thresholds. All features are sorted from largest to smallest by their contribution, and the top five features and their corresponding contribution values ​​are used to form a five-dimensional vector. This vector will be used as one of the inputs for adjusting the grading thresholds to measure the main driving factors of customer policy cancellation risk. If the top five features by contribution include dynamic features constructed from the multi-source data fusion module, such as the cancellation intention intensity index or product attractiveness comparison index, it indicates that external factors or behavioral factors play a dominant role in the customer's policy cancellation decision. In the policy surrender risk prediction module, the calculation process for the market volatility factor is as follows: Based on the time series of the macroeconomic stress index output by the multi-source data fusion module, the macroeconomic stress index values ​​of the most recent four quarters are taken, the first difference between adjacent quarters is calculated, and the first difference is then subtracted again to obtain the second difference. The absolute value of the second difference is taken and the weighted sum is calculated. The result is the market volatility factor. The time series of the macroeconomic stress index is output by the multi-source data fusion module on a quarterly basis, and a value between zero and one is obtained for each quarter. The current quarter and the three quarters before it are taken, for a total of four values, which are denoted as the first quarter value, the second quarter value, the third quarter value, and the fourth quarter value, with the fourth quarter value being the latest value. The calculation process of the first-order difference is as follows: subtract the first-quarter value from the second-quarter value to obtain the first-order difference, subtract the second-quarter value from the third-quarter value to obtain the second-order difference, and subtract the third-quarter value from the fourth-quarter value to obtain the third-order difference. The calculation process for the second-order difference is as follows: subtract the first-order difference from the second-order difference to obtain the first-order difference, and subtract the second-order difference from the third-order difference to obtain the second-order difference; take the absolute values ​​of the two second-order differences, and then calculate the arithmetic mean, that is, add the two absolute values ​​and divide by two to obtain the market volatility factor; the value of this factor ranges from zero to two, where zero indicates that the macroeconomic pressure is stable and unchanged, and two indicates that the pressure fluctuates sharply between adjacent quarters; A higher market volatility factor value indicates that macroeconomic pressures have fluctuated sharply in the near term, and market uncertainty has increased. When the market volatility factor is greater than 0.5, it is considered that the macroeconomic environment has fluctuated significantly. At this time, customers are more sensitive to the value of their policies, and the risk of surrender may be concentrated due to external shocks. This factor will be used to dynamically adjust the classification threshold, so that the system can identify high-risk customers earlier during periods of market turmoil. In the policy surrender risk prediction module, the dynamic adjustment process of the tiered thresholds is as follows: Based on the average contribution of the top five features and the market volatility factor, the three preset basic risk thresholds—extremely high risk, high risk, and medium risk—are adjusted. The average contribution of the top five features is equal to the sum of their contributions divided by five. The preset values ​​for the basic risk thresholds are: 0.80 for extremely high risk, 0.55 for high risk, and 0.30 for medium risk. These three thresholds are initial values ​​obtained through validation set optimization during model training and are used to classify customer policy surrender probabilities into four risk levels. The calculation process for the average contribution of the top five features is as follows: the contribution values ​​of the five features selected in step four are added together and then divided by five to obtain a value between zero and one. This average reflects the comprehensive importance of the main features affecting the customer's policy surrender risk; a higher average indicates that the customer's surrender decision is driven by a few key factors, requiring more refined threshold adjustments. The correction method is as follows: subtract the product of the preset market volatility adjustment coefficient and the market volatility factor mapped by the hyperbolic tangent function from each basic grading threshold, and then subtract the product of the preset feature contribution adjustment coefficient and the average feature contribution. The adjusted grading threshold is obtained by subtracting the product of the preset market volatility adjustment coefficient and the feature contribution average. The preset market volatility adjustment coefficient is 0.05, and the preset feature contribution adjustment coefficient is 0.03. The hyperbolic tangent function maps the market volatility factor to between zero and one. The mapping process is as follows: first calculate the difference between e raised to the power of twice the market volatility factor and one, and divide it by the sum of e raised to the power of twice the market volatility factor and one to obtain the mapping value. This mapping value increases with the increase of the market volatility factor, but the growth rate gradually slows down to avoid excessive downward adjustment of the threshold. The adjusted threshold calculation process is as follows: Taking the extremely high risk threshold as an example, the adjusted extremely high risk threshold is equal to 0.80 minus 0.05 multiplied by the mapping value, and then minus 0.03 multiplied by the average feature contribution. The calculation methods for the high risk threshold and the medium risk threshold are the same, only the base threshold is different. The lower limit of the adjusted threshold is not lower than 80% of the base threshold, that is, the extremely high risk threshold is not lower than 0.64, the high risk threshold is not lower than 0.44, and the medium risk threshold is not lower than 0.24, to prevent the threshold from being too low and causing a large number of normal customers to be misjudged as high risk. The adjusted risk threshold is no less than 80% of the basic risk threshold to achieve dynamic and accurate risk classification of customers. The risk classification rules are as follows: the customer's surrender probability is compared with the three adjusted thresholds in sequence; if the surrender probability is greater than or equal to the adjusted extremely high risk threshold, the customer is marked as extremely high risk; if the surrender probability is greater than or equal to the adjusted high risk threshold but less than the adjusted extremely high risk threshold, the customer is marked as high risk; if the surrender probability is greater than or equal to the adjusted medium risk threshold but less than the adjusted high risk threshold, the customer is marked as medium risk; if the surrender probability is less than the adjusted medium risk threshold, the customer is marked as low risk. Each record in the generated customer list contains a unique customer identifier, surrender probability value, risk level label, and the names and corresponding contribution values ​​of the top five features in terms of feature contribution. This list will be directly called by the retention intervention module to match personalized retention intervention strategies.

[0023] In this embodiment, it is specifically necessary to explain that the process of the strategy engine matching personalized retention intervention strategies in the retention intervention module is as follows: For each customer, the strategy engine iterates through all its built-in strategy templates and calculates the matching score for each template. The strategy engine's built-in strategy templates include three categories: premium reduction strategy templates, coverage upgrade strategy templates, and exclusive service strategy templates. Each strategy template is preset with an optimal applicable risk level: the optimal applicable risk level for the premium reduction strategy template is extremely high risk, the optimal applicable risk level for the coverage upgrade strategy template is high risk, and the optimal applicable risk level for the exclusive service strategy template is medium risk. The matching score is calculated using a weighted summation method, with the following weights for the three sub-items: risk suitability weight is 0.4, historical preference similarity weight is 0.3, and real-time urgency weight is 0.3, with the sum of the three being 1. The matching score is obtained by weighted summation of three sub-items: risk suitability, historical preference similarity, and real-time urgency. Risk fit measures the degree of match between a strategy and a client's risk level. It is calculated by mapping risk levels to numerical values ​​and then calculating the absolute difference between this value and the optimal applicable risk level preset in the strategy template. The rules for mapping risk levels to numerical values ​​are: extremely high risk is mapped to 1.0, high risk to 0.8, medium risk to 0.5, and low risk to 0.2. The calculation process for risk fit is as follows: first, obtain the client's risk level mapping value; then, obtain the mapping value of the optimal applicable risk level preset in the strategy template; calculate the absolute value of the difference between the two mapping values; and then subtract this absolute value from 1 to obtain the risk fit. This value ranges from 0 to 1; a larger value indicates a better match between the strategy and the client's risk level. Historical preference similarity is measured by the cosine similarity between the historical preference vector representing the customer's historical preferences and the corresponding strategy feature vector of the strategy template. The customer's historical preference vector consists of three components: the first component is the percentage of times the customer has accepted the premium reduction strategy in the past, the second component is the percentage of times the customer has accepted the protection upgrade strategy in the past, and the third component is the percentage of times the customer has accepted the exclusive service strategy in the past. The sum of the three components is 1. The strategy feature vector is also a three-dimensional vector, and each strategy template corresponds to a fixed strategy feature vector: the strategy feature vector of the premium reduction strategy template is (1.0, 0.0, 0.0), the strategy feature vector of the protection upgrade strategy template is (0.0, 1.0, 0.0), and the strategy feature vector of the exclusive service strategy template is (0.0, 0.0, 1.0). The calculation process of cosine similarity is as follows: perform a dot product operation between the customer's historical preference vector and the strategy feature vector, that is, multiply the corresponding components and sum them, and then divide by the product of the magnitudes of the two vectors; the magnitude is calculated as the square root of the sum of the squares of each component; the cosine similarity value ranges from -1 to 1, and the larger the value, the more the customer's historical preferences match the strategy type. Real-time urgency is calculated based on the customer's policy cancellation probability and by representing real-time behavior as a real-time behavior vector through a non-linear combination function. The real-time behavior vector consists of two components: the first component is the emotional polarity score of the most recent customer service consultation, with a value range of -1 to 1, where negative values ​​indicate negative emotions and positive values ​​indicate positive emotions; the second component is the current online status flag, with a value of 0 or 1, where 1 indicates that the customer is currently reachable. The calculation process of the nonlinear combination function is as follows: First, the emotional polarity score is converted into an emotional negativity index. The conversion method is to subtract the emotional polarity score from 1 and then divide by 2 to obtain a value ranging from 0 to 1. The larger the value, the more negative the emotion. Then, the probability of withdrawal is calculated by multiplying it by 1 and adding it to the emotional negativity index, and then multiplying it by 1 and adding it to the product of 0.5 times the online status flag to obtain the median value. Finally, the hyperbolic tangent function is applied to this median value. The hyperbolic tangent function is calculated as follows: the difference between twice the power of the median value of e and 1, divided by twice the power of the median value of e and 1; the result is the real-time urgency, ranging from 0 to 1, with a larger value indicating a higher degree of urgency for intervention. The strategy with the highest matching score is selected as the optimal intervention strategy for the client. The matching score is calculated in the following order: first, calculate the risk fit, then calculate the historical preference similarity, and finally calculate the real-time urgency. The scores are then weighted and summed to obtain the total score. The above calculation is repeated for all strategy templates to obtain three matching scores. The strategy template with the highest score is selected as the optimal intervention strategy. If multiple strategy templates have the same score and are all the highest scores, the strategy template with higher risk fit is selected first. If the risk fit is also the same, the strategy template with higher real-time urgency is selected first. In the retention intervention module, the process of pushing intervention commands through the multi-channel distribution engine is as follows: After selecting the optimal intervention strategy, the strategy engine instantiates specific intervention content based on the strategy template's parameter rules and the customer's specific information. This includes reduction ratios, coverage extensions, or exclusive service content. The parameter rules for the strategy template are defined as follows: For the premium reduction strategy template, the premium amount is taken from the static attributes of the policy information in the wide table dataset, i.e., the amount of premium payable per period agreed upon by the customer at the time of insurance purchase, in yuan. The reduction ratio is based on a 5 percentage point base, increasing by 5 percentage points for every 0.1% increase in the customer's surrender probability, and then adding an additional percentage determined by the premium amount tier: 0% for premiums less than 10,000 yuan, 3% for premiums between 10,000 and 50,000 yuan, and 6% for premiums greater than 50,000 yuan; the final reduction ratio will not exceed 30 percentage points. For the coverage upgrade strategy template, extensions are recommended based on the customer's existing coverage gaps, prioritizing coverages related to the customer's most recent customer service consultation. For the exclusive service strategy template, the parameter rules are as follows: Customers are assigned dedicated customer service levels based on their historical service response rate. Those with a response rate higher than 0.7 are assigned to the VIP customer service channel, while those with a response rate lower than 0.7 are assigned to regular dedicated customer service. The premium amount involved in the premium reduction strategy template is taken from the static attributes of the policy information in the wide table dataset, i.e., the amount of premium payable per period agreed upon by the customer when purchasing the policy, in yuan. The calculation process for the reduction ratio is as follows: the basic reduction ratio is five percentage points, plus an increment of five percentage points corresponding to every 0.1% increase in the customer's surrender probability, plus an additional ratio determined according to the premium amount tier—the additional ratio is 0% for premiums less than 10,000 yuan, 3% for premiums between 10,000 and 50,000 yuan, and 6% for premiums greater than 50,000 yuan; the sum of the three constitutes the final reduction ratio, and the final reduction ratio shall not exceed thirty percentage points. The instantiation process is as follows: the strategy engine reads information such as the customer's surrender probability, premium amount, existing coverage, and historical service response rate, substitutes them into the parameter rules to calculate specific values, and generates intervention instructions containing specific parameters. The instantiated intervention command is simultaneously pushed to the SMS gateway, APP push service, and customer service outbound call system through a multi-channel distribution engine. The multi-channel distribution engine maintains a priority list of push channels, with priority from high to low as follows: APP push, SMS push, and customer service outbound call. Synchronous push means that all channels receive the same intervention command at the same time, but the push content format of different channels is different: APP push uses rich text card format, including title, body and jump link; SMS push uses plain text format, with a character limit of 70 characters; customer service outbound call uses speech synthesis script, which is read by agents or an automated speech system. The selection of push channels is based on the online status flag in real-time behavior. If the customer is currently online, push notifications are prioritized through the app; if offline, they are reached through SMS and customer service outbound calls in parallel. The online status flag is obtained as follows: the client app sends a heartbeat packet to the server every 5 minutes. If the most recent heartbeat packet is received within 10 minutes of the current time, the online status flag is 1; otherwise, it is 0. When the online status flag is 1, intervention instructions are sent only through the app push channel, without triggering SMS and customer service outbound calls to avoid disturbing the customer. When the online status flag is 0, intervention instructions are sent simultaneously through the SMS gateway and the customer service outbound call system to ensure that the customer can receive the intervention content through at least one channel. All push operations are completed within a preset time and an intervention log is recorded. The intervention log includes the customer identifier, intervention timestamp, push channel, strategy type, strategy instantiation parameters, customer surrender probability, risk level, matching score, and the customer's subsequent feedback behavior. The preset time is set to 24 hours, meaning that the time interval from the generation of the intervention instruction by the strategy engine to the completion of pushes across all channels does not exceed 24 hours. The intervention log is recorded in structured JSON format, and each log entry includes the following fields: unique customer identifier, intervention trigger time (accurate to the second), push channel name (e.g., "APP push," "SMS," "customer service outbound call"), and strategy type. The system includes the following parameters: policy type (e.g., "premium reduction", "protection upgrade", "exclusive service"), strategy instantiation parameters (e.g., reduction percentage, protection extension name, exclusive service level), customer surrender probability (two decimal places), risk level labels (e.g., "extremely high", "high", "medium", "low"), matching score (three decimal places), and customer follow-up feedback behavior (including whether they opened push notifications, clicked links, initiated inquiries, withdrew surrender applications, made timely payments, increased policy limits, etc., listed as a comma-separated enumeration list). Intervention logs are stored in the database for the closed-loop optimization module to perform correlation analysis based on customer identifiers and time ranges.

[0024] In this embodiment, it is specifically noted that in the closed-loop optimization module, the intervention log is linked with the customer surrender probability, product attractiveness comparison index, and aligned macroeconomic data using the customer identifier and timestamp as keys, constructing an extended dataset containing pre-intervention characteristics and post-intervention behavioral results. The specific process of the association operation is as follows: using the customer's unique identifier and intervention timestamp as the joint key, each record in the intervention log table is left-joined with the customer surrender probability table output by the surrender risk prediction module, the product attractiveness comparison index table output by the multi-source data fusion module, and the aligned macroeconomic data table; Each record in the extended dataset contains the following fields: customer identifier, intervention timestamp, customer surrender probability value before intervention (rounded to four decimal places), product attractiveness comparison index value before intervention (rounded to four decimal places), macroeconomic pressure index value before intervention (rounded to four decimal places), intervention strategy type, strategy instantiation parameters, whether the surrender application was withdrawn after intervention (value is 0 or 1), whether the payment was made on time after intervention (value is 0 or 1), and whether the policy limit was increased after intervention (value is 0 or 1). This extended dataset is stored in a relational database for subsequent retention conversion index calculation and model updates. The process for determining the retention conversion index is as follows: Three behavioral outcomes following customer intervention—whether the cancellation request is withdrawn, whether premiums are paid on time, and whether the policy limit is increased—are assigned different baseline weights and elasticity coefficients. The contribution value of each outcome is calculated using a non-linear exponential function, and the three contribution values ​​are then summed to obtain the retention conversion index. The withdrawal of the cancellation request serves as the decisive signal for the success of the intervention; if the customer withdraws the cancellation request, the intervention is directly considered successful, regardless of the magnitude of the retention conversion index. Only when the customer does not withdraw the cancellation request is the success of the intervention determined based on whether the retention conversion index reaches a preset threshold. The specific form of the nonlinear exponential function is as follows: the contribution value of each behavioral outcome is equal to the baseline weight of that behavior multiplied by the negative elasticity coefficient (in parentheses minus the natural constant) multiplied by the difference of the power of the behavioral outcome value. The baseline weights for the three behavioral outcomes are: 0.5 for withdrawing a policy cancellation application, 0.3 for timely payment, and 0.2 for increasing the policy limit. The three elasticity coefficients are: 2.0 for withdrawing a policy cancellation application, 1.5 for timely payment, and 1.0 for increasing the policy limit. Taking withdrawing a policy cancellation application as an example, if the customer withdraws the cancellation application, the behavioral outcome value is 1, and the contribution value is equal to 0.5 multiplied by the negative elasticity coefficient (in parentheses minus the natural constant). The difference between -2.0 and 1 raised to the power of the natural constant is approximately 0.432. If the customer does not withdraw the cancellation application, the behavior result is 0, and the contribution value is 0. The sum of the three contribution values ​​yields the retention conversion index, which ranges from 0 to 0.791. A larger value indicates a better intervention effect. When the customer withdraws the cancellation application, regardless of the retention conversion index value, the intervention is directly judged as successful. This judgment has higher priority than the judgment based on the retention conversion index threshold. The elasticity coefficient controls the saturation rate of the contribution value. The larger the elasticity coefficient, the faster the contribution value approaches the benchmark weight when the behavior result is 1, avoiding excessive impact of a single behavior on the index. The baseline weights are set as follows: withdrawing a policy cancellation application is the most direct evidence of customer retention, therefore it is assigned the highest weight of 0.5; timely payment indicates that the customer is willing to continue to fulfill their contractual obligations, which is a positive signal for retention, and is assigned a medium weight of 0.3; increasing the policy limit is an additional value-added behavior, which is positive but relatively rare, and is assigned a low weight of 0.2; the elasticity coefficient is set so that when the behavior occurs, the contribution value approaches the baseline weight in a non-linear way. For example, the elasticity coefficient of 2.0 for withdrawing a policy cancellation application means that a single withdrawal can reach 86.5% of the baseline weight, while the elasticity coefficient of 1.0 for increasing the policy limit means that a single increase can only reach 63.2% of the baseline weight. This design ensures the differentiation of different behaviors, while avoiding the excessive dominance of the indicator by a single behavior. In the closed-loop optimization module, the process of updating the policy surrender risk prediction model is as follows: The diagonal elements of the Fisher information matrix for each parameter in the current surrender risk prediction model are extracted. These diagonal elements measure the importance of each parameter to historical data. The calculation of the diagonal elements of the Fisher information matrix uses a Monte Carlo approximation method. The specific steps are as follows: First, 1000 samples are randomly selected from the historical training set of the surrender risk prediction model. For each sample, the partial derivative of the log-likelihood function with respect to each parameter of the model is calculated, and the square of this partial derivative is taken as the sample's contribution to that parameter. The arithmetic mean of the contributions of all 1000 samples is taken to obtain the diagonal elements of the Fisher information matrix for each parameter. The surrender risk prediction model is a gradient boosting decision tree, whose parameters include the splitting feature index, splitting threshold, and output value of each leaf node for each decision tree. The larger the diagonal element of the Fisher information matrix, the more critical the parameter's fit to historical data, and the stronger the constraints it should be in subsequent updates to prevent catastrophic forgetting. Then, a loss function incorporating the new intervention data is constructed. This loss function consists of the logarithmic loss of the new intervention data and an elasticity penalty term based on the diagonal elements of the Fisher information matrix. The elasticity penalty term is the sum of the diagonal elements of the Fisher information matrix multiplied by the square of the difference between the old and new parameters, and then multiplied by the regularization strength. The new intervention data comes from records of successful interventions in the extended dataset. The rule for determining successful intervention is: first, determine whether the customer has withdrawn their cancellation request; if so, the intervention is directly considered successful. If the cancellation request has not been withdrawn, then determine success based on whether the retention conversion index reaches a preset threshold of 0.45. If the retention conversion index is greater than or equal to 0.45, the intervention is considered successful; otherwise, it is considered a failure. For each new intervention data point, the intervention... The model uses the customer policy cancellation probability, product attractiveness comparison index, macroeconomic pressure index, and original wide table features as input features, and the success of the intervention as a binary label (0 for success, 1 for failure) to calculate the log loss. The calculation process of the elasticity penalty term is as follows: iterate through all parameters of the model, for each parameter, take the value of the diagonal element of the Fisher information matrix, multiply it by the square of the difference between the updated value and the original value of the parameter, and then multiply it by the regularization strength of 0.1. Sum the values ​​of this term for all parameters. The total value of the loss function is the sum of the log loss of all new intervention data plus the elasticity penalty term. The design of this loss function allows the model to fit new data while constraining important parameters from large deviations through the penalty term, thereby preserving historical knowledge. The parameters of the policy surrender risk prediction model are incrementally updated by optimizing the loss function. This process absorbs knowledge from new intervention data while retaining knowledge learned from historical data. The optimization process uses the Adam optimizer with a learning rate of 0.0001 and 10 iterations. In each iteration, the gradient of the loss function with respect to all parameters is calculated, and the parameter values ​​are updated according to Adam's adaptive learning rate mechanism. After 10 iterations, the updated parameters of the policy surrender risk prediction model are obtained. This model will be deployed to the policy surrender risk prediction module for predicting the probability of policy surrender in the next cycle. The process of updating the strategy library in the strategy engine is as follows: For each strategy template, it first determines whether the customer has withdrawn their cancellation request. If the customer withdraws the cancellation request, the intervention is directly considered successful. If the customer does not withdraw the cancellation request, the intervention is judged based on whether the retention conversion index reaches the preset threshold of 0.45. If the retention conversion index is greater than or equal to 0.45, the intervention is considered successful; otherwise, it is considered a failure. The posterior success rate distribution of the strategy template is updated based on the success or failure of the intervention, thereby adjusting the weight of the strategy template in subsequent matching to achieve adaptive evolution of the strategy library. Each strategy template maintains a Beta distribution as its posterior success rate distribution, with an initial prior of Beta(1,1), meaning the number of successes and failures are both 1. After each intervention, the root... The parameters of the Beta distribution are updated based on the judgment results: for several pre-successful attempts, the success count is incremented by 1; for several pre-failure attempts, the failure count is incremented by 1; the posterior expected success rate is equal to the number of successes divided by the sum of the number of successes and failures; when adjusting the weight of the strategy template in subsequent matching, the upper confidence bound method is used to calculate the weight adjustment factor, which is equal to the posterior expected success rate plus the exploration coefficient multiplied by the natural logarithm of the total number of interventions (in parentheses) divided by the square root of the quotient of the sum of the number of successes and failures, with the exploration coefficient set to 0.2; this weight adjustment factor will act as an additional multiplier on the real-time urgency sub-item of the matching score in the next cycle of strategy matching, so that strategy templates with higher success rates or fewer exploration attempts will receive higher push priority.

[0025] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0026] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0027] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0028] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0029] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0030] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0031] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the invention's technical solutions and their equivalents, the invention also intends to include these modifications and variations.

Claims

1. An intelligent system for predicting and intervening in the risk of insurance policy surrender, characterized in that: Specifically, it includes: Multi-source data fusion module: Integrates insurance customer policy information, customer behavior data, and external data. Customer behavior data includes policy inquiry frequency, customer service consultation content, and payment timeliness. External data includes industry competitor information and macroeconomic data. The module performs cleaning, alignment, and normalization through an extraction, transformation, and loading pipeline. It uses natural language processing to extract keywords indicating policy cancellation intentions from customer service consultation content, transforms competitor information into product attractiveness comparison indicators, and aligns macroeconomic data by timestamp to construct a wide-table dataset containing static attributes and dynamic time-series features. The policy surrender risk prediction module uses a gradient boosting decision tree to train a policy surrender risk prediction model based on a wide table dataset, and outputs the probability of each customer surrendering their policy within a future time window. The feature contribution degree is calculated based on the contribution of each feature in the wide table dataset to the model output, and the market volatility factor is calculated based on macroeconomic data. The risk classification threshold is dynamically adjusted based on the contribution of features and market volatility factors to accurately classify customers and generate a customer list with risk level labels and customer surrender probability. Retention Intervention Module: Based on the risk level and customer surrender probability in the customer list, combined with the customer's historical preferences and real-time behavior extracted from the wide table dataset, the module matches personalized retention intervention strategies through the strategy engine, pushes intervention instructions through the multi-channel distribution engine, and records intervention logs. Closed-loop optimization module: It performs correlation analysis with intervention logs, customer surrender probability, product attractiveness comparison indicators, and aligned macroeconomic data. It determines retention conversion indicators based on whether the customer cancels the surrender application, pays on time, or increases the policy limit after intervention. It updates the surrender risk prediction model and the strategy library in the strategy engine based on the retention conversion indicators to achieve closed-loop optimization.

2. The intelligent system for predicting and intervening in the risk of insurance customer policy surrender as described in claim 1, characterized in that: In the multi-source data fusion module, the feature construction for customer behavior data specifically includes: To determine the frequency of policy inquiries, an exponentially weighted moving average method was used to calculate the short-term and long-term averages, respectively. The difference between the two was used as an indicator of the change in inquiry activity. A positive value of the indicator indicates that the inquiry frequency has increased recently, while a negative value indicates that it has decreased. For customer service inquiries, a pre-trained language model is used to encode them into semantic vectors. The cosine similarity between the vectors and the semantic center vectors of a pre-built keyword library for policy cancellation is calculated. The vectors are then mapped to a cancellation intention strength index between zero and one through a Logistic transformation. The closer the cancellation intention strength index is to one, the stronger the intention to cancel the policy. To address the timeliness of payments, the non-negative portion of the actual number of days of payment delay is nonlinearly compressed using a hyperbolic tangent function to obtain a delay severity index with a value ranging from zero to one. The delay severity index tends to one as the number of days of delay increases.

3. The intelligent system for predicting and intervening in the risk of insurance customer policy surrender as described in claim 2, characterized in that: In the multi-source data fusion module, the feature construction of external data specifically includes: Based on information about industry competitors, we compare our own products with all competitors one by one, calculate the logarithm of the premium ratio, the difference in coverage, and the weighted Euclidean distance of the promotional efforts, take the minimum value, and then normalize it using the Logistic function to obtain a product attractiveness comparison index with a value between zero and one. The larger the product attractiveness comparison index value, the more obvious the overall attractiveness disadvantage of our own products relative to the competitors corresponding to the minimum value. For the aligned macroeconomic data, after normalizing the range of GDP growth rate, disposable income growth rate and benchmark interest rate, the macroeconomic pressure index is obtained by combining the elasticity coefficients and substituting them into the Logistic function. The macroeconomic pressure index is between zero and one. The closer the macroeconomic pressure index is to one, the greater the macroeconomic pressure.

4. The intelligent system for predicting and intervening in the risk of insurance customer policy surrender as described in claim 3, characterized in that: In the policy surrender risk prediction module, the policy surrender risk prediction model is trained using a gradient boosting decision tree based on a wide table dataset, and is used to output the probability of each customer surrendering their policy within a future time window. The calculation process of feature contribution is as follows: For each feature in the wide table dataset, multiple random perturbations are performed within the value range of each feature in the wide table dataset. During each perturbation, the feature value is replaced with a value randomly sampled from the feature's empirical distribution, while keeping other features unchanged. The feature is then re-input into the surrender risk prediction model to calculate the surrender probability and obtain the perturbed probability sequence. The feature contribution is equal to the average of the absolute values ​​of the difference between the probability of policy cancellation before and after the disturbance. The larger the average value, the more significant the influence of the feature on the probability of policy cancellation. The top five features in terms of contribution and their contribution values ​​are selected for dynamic adjustment of subsequent classification thresholds.

5. The intelligent system for predicting and intervening in the risk of insurance customer policy surrender as described in claim 4, characterized in that: In the policy surrender risk prediction module, the calculation process for the market volatility factor is as follows: Based on the time series of the macroeconomic pressure index output by the multi-source data fusion module, the macroeconomic pressure index values ​​of the most recent four quarters are taken, the first difference between adjacent quarters is calculated, and the first difference is then subtracted again to obtain the second difference. The absolute value of the second difference is taken and the weighted sum is calculated. The result is the market volatility factor. A higher market volatility factor value indicates that macroeconomic pressures have fluctuated sharply in the near term, and market uncertainty has increased.

6. The intelligent system for predicting and intervening in the risk of insurance customer policy surrender as described in claim 5, characterized in that: In the policy surrender risk prediction module, the dynamic adjustment process of the tiered thresholds is as follows: Based on the average contribution of the top five features and the market volatility factor, the three basic classification thresholds of extremely high risk, high risk and medium risk are modified; whereby the average contribution of the top five features is equal to the sum of the contributions of the top five features divided by five. The correction method is as follows: subtract the product of the preset market volatility adjustment coefficient and the market volatility factor mapped by the hyperbolic tangent function from each basic grading threshold, and then subtract the product of the preset feature contribution adjustment coefficient and the mean feature contribution to obtain the adjusted grading threshold. The adjusted lower limit of the classification threshold shall not be lower than 80% of the basic classification threshold, so as to achieve dynamic and accurate classification of customer risk levels.

7. The intelligent system for predicting and intervening in the risk of insurance customer policy surrender as described in claim 6, characterized in that: In the retention intervention module, the process by which the strategy engine matches personalized retention intervention strategies is as follows: For each customer, the strategy engine iterates through all its built-in strategy templates and calculates the matching score of each strategy template for the customer. The matching score is obtained by weighted summation of three sub-items: risk suitability, historical preference similarity, and real-time urgency. Risk fit measures the degree of match between the strategy and the client's risk level. It is calculated by mapping the risk level to a numerical value and then calculating the absolute difference between the value and the optimal applicable risk level preset in the strategy template. Historical preference similarity measures the cosine similarity between the customer's historical preferences, represented as a historical preference vector, and the policy feature vector corresponding to the policy template. Real-time urgency is calculated based on the customer's probability of canceling the policy and by representing real-time behavior as a real-time behavior vector through a nonlinear combination function. The strategy with the highest matching score is selected as the optimal intervention strategy for the client.

8. The intelligent system for predicting and intervening in the risk of insurance customer policy surrender as described in claim 7, characterized in that: In the retention intervention module, the process of pushing intervention commands through the multi-channel distribution engine is as follows: After selecting the optimal intervention strategy, the strategy engine instantiates specific intervention content based on the parameter rules of the strategy template and the customer's specific information, including the reduction ratio, extended protection items, or exclusive service content. The instantiated intervention instructions are simultaneously pushed to the SMS gateway, APP push service and customer service outbound call system through a multi-channel distribution engine; The selection of push channels is based on the online status indicator in real-time behavior. If the customer is currently online, push will be sent via the APP first. If the customer is offline, they will be reached through SMS and customer service outbound calls in parallel. All push operations are completed within a preset time and an intervention log is recorded. The intervention log includes customer identifier, intervention timestamp, push channel, strategy type, strategy instantiation parameters, customer surrender probability, risk level, matching score, and customer's subsequent feedback behavior.

9. The intelligent system for predicting and intervening in the risk of insurance customer policy surrender as described in claim 8, characterized in that: In the closed-loop optimization module, the intervention logs are linked with customer surrender probability, product attractiveness comparison index, and aligned macroeconomic data using customer identifier and timestamp as keys to construct an extended dataset containing pre-intervention characteristics and post-intervention behavioral results. The process of determining the retention conversion index is as follows: assign different benchmark weights and elasticity coefficients to the three behavioral outcomes after customer intervention, namely whether the cancellation application is withdrawn, whether the premium is paid on time, and whether the policy limit is increased. Calculate the contribution value of each behavioral outcome through a nonlinear exponential function, and then sum the three contribution values ​​to obtain the retention conversion index. Among them, the benchmark weight for withdrawing the policy cancellation application is the highest, followed by timely payment, and the weight for increasing the policy amount is the lowest. The elasticity coefficient controls the saturation rate of the contribution of each behavior to the indicator, so that the marginal contribution of a single behavior to the indicator decreases as the number of times the behavior occurs increases.

10. The intelligent system for predicting and intervening in the risk of insurance customer policy surrender as described in claim 9, characterized in that: In the closed-loop optimization module, the process of updating the policy surrender risk prediction model is as follows: Extract the diagonal elements of the Fisher information matrix for each parameter from the current surrender risk prediction model. The diagonal elements measure the importance of each parameter to historical data. Then, a loss function containing the new intervention data is constructed. The loss function consists of the log loss of the new intervention data and the elasticity penalty term based on the diagonal elements of the Fisher information matrix. The elasticity penalty term is the sum of the diagonal elements of the Fisher information matrix multiplied by the square of the difference between the old and new parameters and then multiplied by the regularization strength. By optimizing the loss function, the parameters of the surrender risk prediction model are updated incrementally, absorbing the knowledge contained in the new intervention data while retaining the knowledge learned from the historical data. The process of updating the strategy library in the strategy engine is as follows: For each strategy template, it is first determined whether the customer has withdrawn the cancellation application. If the cancellation application is withdrawn, the intervention is directly determined to be successful. If the cancellation application is not withdrawn, the intervention is determined to be successful based on whether the retention conversion index exceeds the preset threshold. The posterior success rate distribution of the strategy template is updated based on the success or failure of the intervention, and then the weight of the strategy template in subsequent matching is adjusted to achieve adaptive evolution of the strategy library.

Images