Insurance agent three-turn rate improvement algorithm and system

Abstract

The invention belongs to the technical field of insurance digital services, and particularly discloses an insurance agent three-transformation rate improvement algorithm and system, and the algorithm comprises five core steps: data preprocessing, feature extraction, three-transformation potential prediction, personalized intervention strategy generation and strategy iterative optimization. According to the method, basic information, training data, exhibition data and customer interaction data of insurance agents are fused, an integrated learning model is adopted to accurately predict the three-transition probability of the agents, differentiation intervention schemes are generated in combination with agent portraits, and algorithm parameters are continuously optimized based on real-time feedback data; the system corresponds to an algorithm process and comprises a data acquisition module, a data preprocessing module, an algorithm processing module, an intervention execution module, an effect feedback module and a data storage module, and automation, intelligence and precision of the whole process of three-conversion rate improvement are achieved.

Description

Technical Field

[0001] This invention relates to the fields of digital insurance services and artificial intelligence technology, specifically to an algorithm and system for improving the three-way conversion rate of insurance agents. Background Technology

[0002] As the insurance industry shifts from a "mass recruitment" model to a "lean and professional" one, the retention and development of insurance agents have become a crucial component of insurance companies' core competitiveness. The "three-stage conversion rate" (transition from probationary to professional) is a key indicator of an insurance agent's growth quality, directly reflecting the efficiency of an agent's transition from the probationary to the professional stage. Its level not only affects an insurance company's labor cost control but also determines the stability and business output of the agent force. Currently, the three-stage conversion rate for insurance agents in my country is generally low, with most insurance companies having a rate below 50%. Simultaneously, they face problems such as high agent attrition rates, high training costs, and homogenized intervention measures, severely hindering the high-quality development of the agent force.

[0003] Current methods for improving the conversion rate of insurance agents mainly rely on manual management and experience-based judgment. Managers develop uniform training plans and incentive policies based on agents' basic performance data, lacking precise analysis of individual agent differences. This makes it impossible to identify the key weaknesses affecting each agent's conversion rate, resulting in weakly targeted intervention measures and poor implementation effects. Although some insurance companies have introduced digital tools and simple data statistical analysis methods in recent years, these tools mainly focus on improving agents' customer acquisition efficiency and business development convenience, without building specific algorithmic models for the conversion rate. This makes it impossible to accurately predict the conversion potential and automatically generate personalized intervention strategies.

[0004] Furthermore, existing technologies largely rely on traditional manual operation modes, lacking data-driven intelligent decision support. Therefore, current technologies cannot meet the needs of insurance companies for accurately improving agent conversion rates. There is an urgent need for a technical solution that can integrate multi-dimensional data, use intelligent algorithms to predict conversion potential, provide personalized intervention, and dynamically optimize the process, thereby solving the current technical challenges of low efficiency and poor adaptability in improving conversion rates. Summary of the Invention

[0005] The purpose of this invention is to provide an algorithm and system for improving the three-level conversion rate of insurance agents, solving the technical problems of low efficiency and poor adaptability in improving the three-level conversion rate of insurance agents due to the reliance on human experience, insufficient data-driven approach, and homogenization of intervention measures in existing technologies. This invention achieves automation, intelligence, and precision in the entire process of improving the three-level conversion rate, reduces training costs, and promotes the professional transformation of the agent force.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an algorithm for improving the three-tier conversion rate of insurance agents, the algorithm comprising the following steps: S1. Data Preprocessing: Collect multi-dimensional raw data from insurance agents, and clean, complete, normalize, and remove outliers from the raw data to obtain a standardized dataset; the multi-dimensional raw data includes agent basic information data, training data, business development data, customer interaction data, and assessment data; Data preprocessing is fundamental to the accurate operation of algorithms. Due to inconsistencies in data format, missing data, duplication, and anomalies across different insurance company systems, preprocessing is necessary to eliminate data noise and ensure data quality. Specifically, Lagrange interpolation is used to complete missing data, avoiding feature extraction bias caused by missing data; the 3σ principle is used to identify outliers (such as extreme performance data or abnormal training durations) and replace them with the median of the corresponding features to prevent outliers from affecting model prediction accuracy; min-max normalization is used to map numerical data to the [0,1] interval, eliminating dimensional differences and allowing data from different dimensions to be directly used for model training; one-hot encoding is used to encode categorical data (such as education level and job title), converting it into a computer-recognizable numerical form, ultimately resulting in a standardized dataset.

[0007] S2. Feature Extraction: Based on a standardized dataset, a combination of mutual information and random forest algorithms is used to extract key features that affect the agent's three-way conversion rate and construct a key feature set. The key features include training completion rate, first-order transaction cycle, customer follow-up frequency, customer satisfaction score, assessment pass rate, and the number of referral leads obtained. Numerous factors influence the conversion rate of insurance agents. Directly inputting all features into the model would increase model complexity, reduce training efficiency, and even lead to overfitting. Therefore, this invention employs a combination of mutual information and random forest algorithms for key feature extraction: First, the mutual information value between each feature and the conversion result (converted / not converted) is calculated using the mutual information method. The larger the mutual information value, the stronger the correlation between the feature and the conversion rate. Initial features with mutual information values ​​greater than a preset threshold (e.g., 0.1) are selected, while features with extremely weak correlations are removed. Second, the initial features are input into the random forest algorithm model, and the importance score of each initial feature is calculated. The feature importance score reflects the degree of influence of the feature on the conversion result. Finally, the top N features are selected as key features based on their importance scores from highest to lowest (N is dynamically adjusted according to the agent sample size; when the sample size is greater than 1000, N is 10-15; when the sample size is less than 1000, N is 5-8), constructing a key feature set. This approach ensures the representativeness of the features while reducing model complexity.

[0008] S3. Third-stage potential prediction: Construct an ensemble learning prediction model, input the key feature set into the ensemble learning prediction model, and output the third-stage probability and non-transformation risk level of each agent; the ensemble learning prediction model is composed of a weighted fusion of a random forest model, an XGBoost model and a LightGBM model, and the model weight coefficients are optimized by a grid search algorithm; Single machine learning models suffer from insufficient prediction accuracy and weak generalization ability in predicting three-way potential. This invention adopts an ensemble learning approach, weighting and fusing multiple single models to fully leverage their respective strengths and improve prediction accuracy. The specific construction process is as follows: First, a random forest model, an XGBoost model, and a LightGBM model are trained separately. The random forest model has advantages such as resistance to overfitting and the ability to handle high-dimensional data, making it suitable for initial feature selection and prediction. The XGBoost and LightGBM models have advantages such as fast training speed and high prediction accuracy, and can capture complex nonlinear relationships between features. A K-fold cross-validation method (K=5) is used to optimize the hyperparameters of each single model (such as the number of decision trees in the random forest and the learning rate in the XGBoost) to ensure optimal prediction performance for each model. Second, based on the prediction accuracy of each single model, a grid search algorithm is used to determine the weight coefficients of each model. The weight coefficients are correlated with the model's prediction accuracy. The prediction accuracy is positively correlated with the weighting coefficient; that is, the higher the prediction accuracy, the larger the weighting coefficient. For example, if the prediction accuracy of the Random Forest model is 82%, the XGBoost model is 88%, and the LightGBM model is 87%, then the corresponding weighting coefficients can be set to 0.25, 0.4, and 0.35, respectively. Finally, the prediction results of each individual model are weighted and summed with the corresponding weighting coefficients to obtain the third-stage conversion probability of each agent (the value ranges from 0 to 1). Based on the third-stage conversion probability, the risk level of non-conversion is divided as follows: a third-stage conversion probability of <30% is high risk (high probability of not being able to complete the third stage conversion), 30% ≤ third-stage conversion probability <70% is medium risk (there is some potential for third-stage conversion, requiring key intervention), and a third-stage conversion probability ≥70% is low risk (high probability of completing the third stage conversion, requiring appropriate guidance).

[0009] S4. Personalized intervention strategy generation: Based on the agent's three-way conversion probability, non-conversion risk level, and key characteristic weaknesses, and combined with a preset intervention strategy library, a targeted personalized intervention strategy is generated; the intervention strategy library includes four types of strategies: training, business development, incentive, and supervision. Agents with different risk levels and weaknesses face significantly different problems during the three-stage transition process, making it impossible to achieve ideal improvement results with homogenized intervention strategies. Therefore, this invention generates personalized intervention strategies based on individual agent differences: First, an intervention strategy library is constructed, containing four core strategies. These include: training strategies (basic sales script training, customer follow-up skills training, referral method training, and advanced professional skills training); sales strategies (intelligent sales tool support, precise customer lead delivery, and sales process optimization guidance); incentive strategies (phased small-amount incentives, advanced performance rewards, honors and recognition, and preferential promotion channels); and supervisory strategies (one-on-one mentoring, phased goal setting, and weekly review guidance). Second, for agents with different risk levels, corresponding strategies are matched based on their key characteristics and weaknesses: For high-risk agents, the core problems are usually weak basic skills and insufficient sales motivation. Therefore, supervisory and basic training strategies are prioritized, with dedicated supervisors providing one-on-one mentoring, supplementing basic sales knowledge and script training, and setting phased small-amount incentive goals (such as achieving...). For agents with medium risk, a reward of 50 yuan is given for each effective customer follow-up, gradually improving their sales capabilities and enthusiasm. For agents with medium risk, their basic capabilities meet the standards, but they have a certain weakness (such as insufficient customer follow-up frequency or low number of referral leads). Therefore, specialized training strategies and sales support strategies are matched, and targeted training is conducted based on the weakness (such as follow-up skills training for insufficient customer follow-up frequency). Intelligent sales tools are provided to help them make up for their weaknesses (such as AI customer needs analysis tools and referral poster generation tools). For agents with low risk, they have strong sales capabilities and enthusiasm, and their core needs are to improve their professional level and customer acquisition capabilities. Therefore, incentive strategies and advanced training strategies are matched, advanced performance incentive awards are set up (such as a reward of 2,000 yuan for completing 3 sales), and advanced professional skills training and team management training are conducted to guide them to actively acquire customers and make referrals, and promote their rapid completion of the three conversions. Finally, the generated personalized intervention strategies are bound to the agent's identification and output to the intervention execution module.

[0010] S5. Strategy Iteration and Optimization: Collect real-time behavioral data and three-way transition results data of agents after receiving intervention, calculate the evaluation index of the implementation effect of intervention strategy, and adjust the parameters of integrated learning prediction model and strategy adaptation rules of intervention strategy library according to the evaluation results to achieve dynamic iterative optimization of algorithm.

[0011] The business environment and individual circumstances of insurance agents are constantly changing. Fixed algorithm models and intervention strategies cannot maintain ideal improvement effects in the long term. Therefore, this invention constructs a closed-loop iterative optimization mechanism to ensure the adaptability and effectiveness of the algorithm: First, it defines the evaluation indicators for the implementation effect of intervention strategies, including the post-intervention three-transformation rate (the percentage of agents who complete the three transformations after receiving intervention), the strategy execution rate (the percentage of agents who actually implement the intervention strategy), the agent behavior improvement rate (the percentage of improvement in key characteristic weaknesses after intervention), and the training cost saving rate (the percentage of reduction in unit agent training costs after adopting this algorithm); Second, it calculates the implementation effect evaluation indicators for each intervention strategy and eliminates strategies that do not meet the standards (such as strategies that improve the post-intervention three-transformation rate by less than 5%). Furthermore, the strategy content is optimized (e.g., if basic training strategies are ineffective, the training course content and format are adjusted); secondly, real-time behavioral data of agents after intervention (e.g., frequency of customer follow-up after intervention, training completion rate) and three-conversion result data (converted / not converted) are added to a standardized dataset, and the ensemble learning prediction model is retrained, adjusting model parameters (e.g., hyperparameters, weight coefficients) and feature weights to improve model prediction accuracy; finally, the strategy adaptation rules of the intervention strategy library are updated to make strategy generation more closely match the actual needs of agents (e.g., if agents in a certain area have difficulty obtaining referral leads, the strategy adaptation rules for agents in that area are optimized, and the matching weight of referral-related strategies is increased), realizing closed-loop iteration of the algorithm and ensuring continuous optimization of the three-conversion rate improvement effect.

[0012] A system for improving the three-way conversion rate of insurance agents is provided. The system is used to implement the above-mentioned algorithm for improving the three-way conversion rate of insurance agents. It includes a data acquisition module, a data preprocessing module, an algorithm processing module, an intervention execution module, an effect feedback module, and a data storage module. The modules work together to achieve automation, intelligence, and precision in the entire process of improving the three-way conversion rate.

[0013] Data acquisition module: Used to collect multi-dimensional raw data from insurance agents, including basic agent information, training data, sales data, customer interaction data, and performance evaluation data; the data acquisition module interfaces with the insurance company's existing core business system, training system, CRM system, and intelligent sales platform to achieve real-time data collection; The data acquisition module is the core of the system's data source, and the dimensions of the data it collects directly affect the algorithm's prediction accuracy and the targeting of intervention strategies. Specifically, the collected data includes: basic agent information (age, education, work experience, date of employment, team, and job level), used to analyze the impact of the agent's basic qualifications on the three-conversion rate; training data (training course completion rate, training assessment scores, training participation duration, number of specialized skills training sessions, and training feedback scores), used to assess the agent's learning ability and professional foundation; sales data (first-sale transaction cycle, cumulative number of policies sold, premium size, number of customer follow-ups, follow-up success rate, number of referral leads obtained, and referral conversion rate), used to analyze the agent's sales ability and enthusiasm; customer interaction data (customer consultation response time, customer satisfaction score, number of customer complaints, policy delivery service quality, and customer return visit completion rate), used to assess the agent's customer service ability, which directly affects customer retention and referrals, and thus the three-conversion rate; and performance evaluation data (monthly / quarterly performance evaluation pass rate, performance growth rate, core indicator completion status, and reward and punishment records), used to evaluate the agent's overall performance. The data acquisition module connects to the insurance company's existing system via an API interface, eliminating the need for manual data entry and enabling real-time data collection and synchronous updates to ensure data timeliness.

[0014] Data preprocessing module: Connected to the data acquisition module, it is used to clean, complete, normalize, remove outliers and encode the acquired raw data, and output a standardized dataset; The data preprocessing module receives the raw data transmitted from the data acquisition module and processes the data according to the preprocessing flow of algorithm step S1: First, it removes duplicate and invalid data (such as blank data and data with incorrect format); second, it uses Lagrange interpolation to complete missing data; third, it uses the 3σ principle to remove outliers and replace them; then, it uses the min-max normalization method to normalize the numerical data; finally, it uses one-hot encoding to encode the categorical data to obtain a standardized dataset, and then transmits the standardized dataset to the algorithm processing module.

[0015] Algorithm processing module: Connected to the data preprocessing module, it is the core module of the system, including feature extraction unit, prediction model unit, policy generation unit and iterative optimization unit; The feature extraction unit is used to receive the standardized dataset, extract key features according to the feature extraction process of algorithm step S2, and use a combination of mutual information method and random forest algorithm to construct a key feature set, and transmit the key feature set to the prediction model unit. The prediction model unit is used to run the ensemble learning prediction model, receive the key feature set transmitted by the feature extraction unit, output the three-transfer probability and non-transfer risk level of each agent according to the prediction process of algorithm step S3, and transmit the prediction results to the strategy generation unit. The strategy generation unit is used to receive the three-transfer probability and non-transfer risk level transmitted by the prediction model unit, identify the agent's characteristic weaknesses by combining the key feature set, generate personalized intervention strategies by combining the intervention strategy library according to the strategy generation process of algorithm step S4, and transmit the intervention strategies to the intervention execution module. The iterative optimization unit is used to receive real-time data (behavioral data after intervention, three-stage result data, and effect evaluation data) transmitted by the effect feedback module, evaluate the effect of the intervention strategy according to the iterative optimization process of algorithm step S5, adjust the parameters and feature weights of the integrated learning prediction model, optimize the strategy adaptation rules of the intervention strategy library, realize the closed-loop iteration of the algorithm, and update the optimized model parameters and strategy adaptation rules to the prediction model unit and the strategy generation unit respectively.

[0016] Intervention Execution Module: Connected to the algorithm processing module, it receives personalized intervention strategies, pushes the strategies to agent terminals, supervisor terminals and relevant management terminals, monitors the strategy execution process, and collects strategy execution data. The intervention execution module receives personalized intervention strategies transmitted by the strategy generation unit and pushes these strategies using various methods based on the strategy type and the recipient: for agent terminals, it pushes training notifications, incentive reminders, and business development guidance through the insurance company's agent app; for supervisor terminals, it pushes supervisory tasks, agent weakness information, and supervisory priorities; and for management terminals, it pushes information such as agent risk level distribution and intervention strategy execution progress. Simultaneously, the intervention execution module collects strategy execution data in real time (such as whether agents have completed training, performed customer follow-up tasks, and whether supervisors have completed support work) and transmits the strategy execution data to the effect feedback module.

[0017] Effect feedback module: Connected to the intervention execution module and the algorithm processing module, it is used to collect real-time behavioral data, three-way result data and strategy execution feedback of agents after receiving intervention, calculate intervention effect evaluation indicators, and feed the relevant data back to the iterative optimization unit of the algorithm processing module; The effect feedback module receives strategy execution data from the intervention execution module, and simultaneously collects real-time behavioral data of agents after intervention (such as customer follow-up frequency, training completion rate, and number of referral leads acquired) and three-conversion result data (converted / not converted). In addition, it receives strategy execution feedback from agents, supervisors, and managers (such as strategy rationality and practicality scores). Based on the above data, it calculates intervention strategy implementation effectiveness evaluation indicators (three-conversion rate after intervention, strategy execution rate, agent behavior improvement rate, and training cost savings rate), and transmits the real-time behavioral data, three-conversion result data, strategy execution feedback, and effectiveness evaluation indicators to the iterative optimization unit of the algorithm processing module to provide data support for algorithm iteration.

[0018] Data storage module: Connected to various modules of the system, it is used to store standardized datasets, key feature sets, prediction model parameters, intervention strategy library, strategy execution data, effect evaluation data and three-transformation result data. It adopts encrypted storage to ensure data security and supports real-time data query and traceability. The data storage module adopts a distributed storage architecture, featuring large capacity, high reliability, and high security, capable of storing all data generated during system operation. Simultaneously, it employs AES encryption to encrypt sensitive data (such as agent personal information, customer information, and performance data) to prevent data leakage; data access permissions are set, allowing personnel in different positions to access data only within their corresponding authorized scope, ensuring data security. Furthermore, the data storage module supports real-time data querying and traceability, allowing managers to query agent three-stage prediction results, intervention strategy implementation status, and effect evaluation data at any time, providing data support for management decision-making.

[0019] Furthermore, the system also includes a visualization module connected to the algorithm processing module and the data storage module. This module displays agent conversion rates, the distribution of non-conversion risk levels, the implementation status of intervention strategies, and effectiveness evaluation results in chart form. It supports multi-dimensional queries and statistics by team, region, and onboarding period, providing decision support for managers. The visualization module uses various chart formats such as bar charts, line charts, pie charts, and heatmaps to intuitively display relevant data. For example, a pie chart can show the percentage of agents with different non-conversion risk levels, a line chart can show the trend of conversion rates before and after intervention, and a heatmap can show the distribution of agent conversion rates in different regions. Managers can quickly grasp the agent conversion situation through the visualization interface and formulate targeted management decisions.

[0020] Furthermore, the system adopts a modular design, supporting seamless integration with existing insurance company information systems without requiring large-scale modifications. It also possesses scalability, allowing for the addition of data dimensions, optimization of algorithm models, and expansion of intervention strategy types to meet the agent management needs of insurance companies of different sizes. For example, smaller insurance companies can simplify some module functions (e.g., eliminating the need for new visualization modules), while larger insurance companies can add data dimensions (e.g., agent compliance data) and expand intervention strategy types (e.g., compliance training strategies), thereby improving the system's adaptability. Compared with the prior art, the beneficial effects of the present invention are: Breaking through the limitations of traditional manual management models, this technology employs multi-dimensional data fusion and intelligent algorithm modeling to accurately predict the potential of insurance agents in their third-stage transformation. The prediction accuracy rate is over 88%, and it can accurately identify agents with different risk levels, providing data support for personalized intervention. This solves the problems of low prediction accuracy and lack of specificity in existing technologies. Based on the agent's risk level and characteristic weaknesses, personalized intervention strategies are generated, which are different from the homogeneous intervention measures of existing technologies. This achieves "one policy for one person" and significantly improves the intervention effect. Practical verification has shown that it can increase the insurance agent's third-party turnover rate by more than 26% and effectively reduce the agent dropout rate. The closed-loop algorithm iteration and optimization mechanism can dynamically adjust the model parameters and intervention strategy adaptation rules based on real-time data after intervention, continuously optimize the three-conversion rate improvement effect, and adapt to the dynamic changes in the business environment and individual status of insurance agents. The system adopts a modular design, supports seamless integration with the existing information systems of insurance companies, and does not require large-scale modifications to existing systems. It has strong adaptability and scalability, and can adapt to the agent management needs of insurance companies of different sizes. At the same time, it ensures data security through encrypted data storage and access control, and complies with the data compliance requirements of the insurance industry. The automation and intelligentization of the entire process for improving the three-way conversion rate (conversion rate, sales conversion rate, and agent turnover rate) reduces manual intervention, lowers the agent training and management costs for insurance companies, achieving cost savings of over 22%, promotes the transformation of the insurance agent force towards specialization and refinement, and enhances the core competitiveness of insurance companies. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the algorithm structure of the present invention; Figure 2 This is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Example 1 Please see Figure 1 This embodiment provides an algorithm for improving the three-tier conversion rate of insurance agents. The algorithm includes the following steps: S1. Data Preprocessing: Collect multi-dimensional raw data from 1,000 probationary agents of an insurance company who have been employed for less than one month. This raw data includes: basic agent information (age, education, work experience, start date, team, job level), training data (training course completion rate, training assessment scores, training participation duration, number of specialized skills training sessions, training feedback scores), sales data (first-sale transaction cycle, cumulative number of policies sold, premium amount, number of customer follow-ups, follow-up success rate, number of referral leads obtained, referral conversion rate), customer interaction data (customer inquiry response time, customer satisfaction score, number of customer complaints, policy delivery service quality, customer follow-up completion rate), and performance evaluation data (monthly performance evaluation). The data includes: compliance rate, performance growth rate, completion status of core indicators, and reward and punishment records. Preprocessing of the raw data involves: using Lagrange interpolation to fill in missing training assessment scores, customer satisfaction ratings, etc.; using the 3σ principle to identify outliers (e.g., an agent's customer follow-up count is 1000 times / month, far exceeding the industry average, and is therefore considered an outlier), and replacing them with the median of the corresponding feature; using min-max normalization to map numerical data (e.g., training participation time, customer follow-up count) to the [0,1] interval; and using one-hot coding to encode categorical data (e.g., education levels are divided into high school and below, junior college, bachelor's degree, master's degree and above, and job levels are divided into probationary junior and probationary intermediate levels) to obtain a standardized dataset.

[0024] S2. Feature Extraction: The mutual information method is used to calculate the mutual information value between each feature in the standardized dataset and the three-way referral result. The preset mutual information value threshold is 0.1, and 12 initial features with mutual information values ​​greater than 0.1 are selected. The 12 initial features are input into the random forest algorithm model, and the importance score of each initial feature is calculated. The features are sorted from high to low, and the top 10 features are selected as key features to construct a key feature set. The 10 key features are: training completion rate (importance score 0.18), first order transaction cycle (importance score 0.16), customer follow-up frequency (importance score 0.14), customer satisfaction score (importance score 0.12), assessment pass rate (importance score 0.10), number of referral leads obtained (importance score 0.08), training assessment results (importance score 0.06), follow-up success rate (importance score 0.05), training feedback score (importance score 0.04), and work experience (importance score 0.03).

[0025] S3. Three-stage potential prediction: Construct an ensemble learning prediction model, which is a weighted fusion of a random forest model, an XGBoost model, and a LightGBM model; use K-fold cross-validation (K=5) to optimize the hyperparameters of each individual model: the number of decision trees in the random forest model is set to 100, and the maximum depth is set to 8; the learning rate of the XGBoost model is set to 0.1, the number of trees is set to 100, and the maximum depth is set to 6; the learning rate of the LightGBM model is set to 0.08, the number of trees is set to 120, and the maximum depth is set to 7; and grid search is used to optimize the hyperparameters of each model. The algorithm determines the weight coefficients of each model. The Random Forest model has a prediction accuracy of 82% and a weight coefficient of 0.25; the XGBoost model has a prediction accuracy of 88% and a weight coefficient of 0.4; and the LightGBM model has a prediction accuracy of 87% and a weight coefficient of 0.35. The key feature set is input into the ensemble learning prediction model, which outputs the probability of each agent's third-level conversion and classifies the risk level of non-conversion: a probability of third-level conversion < 30% is high risk (180 people in total), 30% ≤ probability of third-level conversion < 70% is medium risk (520 people in total), and a probability of third-level conversion ≥ 70% is low risk (300 people in total).

[0026] S4. Personalized Intervention Strategy Generation: An intervention strategy library is built, including four types of strategies: training, sales development, incentives, and supervision. For high-risk agents (180 individuals), supervision and basic training strategies are matched: Dedicated supervisors provide one-on-one support, conducting weekly review sessions; basic sales script training and customer follow-up skills training courses are provided, requiring at least 10 hours of training per month; phased small-amount incentive targets are set, with a reward of 50 yuan for each effective customer follow-up and 100 yuan for each training assessment achieved. For medium-risk agents (520 individuals), specialized training and sales support strategies are matched based on their key characteristics and weaknesses: For agents (180 individuals) with insufficient customer follow-up frequency, specialized customer follow-up skills training is provided, along with intelligent follow-up support. The system includes a reminder tool to set daily follow-up goals (at least 3 effective follow-ups); for agents with low referral lead generation (150 people), it provides specialized training on referral methods, a referral poster generation tool, and referral incentive templates; for agents with low training completion rates (190 people), it optimizes the training course delivery schedule, sets training reminders, and arranges peer learning assistance; for low-risk agents (300 people), it matches incentive strategies with advanced training strategies: it sets advanced performance incentives, awarding 2000 yuan for completing 3 sales within 3 months and 5000 yuan for completing 5 sales; it pushes advanced professional skills training (such as high-end customer service skills and team management basics) to guide them to actively acquire customers and make referrals; it binds the generated personalized intervention strategies with agent identifiers and outputs them to the intervention execution module.

[0027] S5. Strategy Iteration and Optimization: Track the intervention process for 3 months, collect real-time behavioral data, three-way referral results data, and strategy execution feedback of agents after receiving the intervention; calculate the evaluation indicators of the intervention strategy implementation effect: the overall three-way referral rate after the intervention was 68%, an increase of 26% compared with the pre-intervention rate (42%); the strategy execution rate was 85%; the agent behavior improvement rate was 78%; the training cost saving rate was 22%; strategies with unsatisfactory results were removed (such as a basic training course whose training completion rate increased by less than 3% after the intervention, which was removed and the course content was optimized); real-time data after the intervention was added to a standardized dataset, the ensemble learning prediction model was retrained, and the model hyperparameters and weight coefficients were adjusted (the weight coefficient of the XGBoost model was adjusted to 0.42, the LightGBM model to 0.33, and the random forest model to 0.25); the strategy adaptation rules of the intervention strategy library were updated, and the matching weight of the referral strategy related to medium-risk agents was increased to achieve the first iteration optimization of the algorithm; subsequent iterations were performed every 3 months to continuously improve the three-way referral rate improvement effect.

[0028] Example 2 Please see Figure 2This embodiment provides a system for improving the three-tier conversion rate of insurance agents, used to implement the algorithm for improving the three-tier conversion rate of insurance agents described in Embodiment 1. It includes a data acquisition module, a data preprocessing module, an algorithm processing module, an intervention execution module, an effect feedback module, a data storage module, and a visualization module. The specific structure is as follows: Data collection module: It connects with the insurance company's existing core business system, training system, CRM system and intelligent sales platform through API interface to collect multi-dimensional raw data from 1,000 trial agents in real time, including agent basic information data, training data, sales data, customer interaction data and assessment data; the data collection frequency is set to synchronize once a day to ensure the timeliness of the data; Data preprocessing module: Connected to the data acquisition module, it receives the acquired raw data, and performs cleaning, completion, normalization, outlier removal and encoding on the data according to the preprocessing process in step S1 of Example 1, outputting a standardized dataset and transmitting it to the algorithm processing module. The algorithm processing module, developed in Python and deployed on a cloud server, includes a feature extraction unit, a prediction model unit, a strategy generation unit, and an iterative optimization unit. The feature extraction unit uses a combination of mutual information and random forest algorithms to extract key features and construct a key feature set. The prediction model unit runs an ensemble learning prediction model, outputting the agent's three-way conversion probability and non-conversion risk level. The strategy generation unit combines an intervention strategy library to generate personalized intervention strategies. The iterative optimization unit receives data from the effect feedback module to achieve closed-loop iteration of the algorithm. Intervention Execution Module: Deployed on insurance company agent APP, supervisor terminal and management terminal, it receives personalized intervention strategies transmitted by the algorithm processing module and pushes them to the corresponding terminals through APP message push, SMS reminder, background notification and other means; it collects strategy execution data in real time (such as agent training completion status, customer follow-up execution status) and transmits it to the effect feedback module; Effect feedback module: Real-time collection of agent behavior data, three-turn result data and strategy execution feedback after intervention, calculation of intervention effect evaluation indicators (three-turn rate after intervention, strategy execution rate, agent behavior improvement rate, training cost saving rate), and feedback of relevant data to the iterative optimization unit of the algorithm processing module; Data storage module: Adopts a distributed storage architecture and uses AES encryption algorithm to encrypt and store sensitive data; the stored content includes standardized datasets, key feature sets, prediction model parameters, intervention strategy library, strategy execution data, effect evaluation data, and three-stage result data; supports real-time data query and traceability, and sets three levels of access permissions (agents can only query their own data, supervisors can query the data of the agents they are responsible for, and managers can query all data). Visualization module: Developed using the ECharts visualization framework, it connects with the algorithm processing module and data storage module to intuitively display agent conversion rates, non-conversion risk level distribution, intervention strategy implementation status, and effect evaluation results in the form of bar charts, line charts, pie charts, heatmaps, etc. It supports multi-dimensional queries and statistics by team, region, and onboarding cycle. Managers can quickly grasp relevant information and make management decisions through the visual interface.

[0029] The system operation process in this embodiment is as follows: The data acquisition module collects raw agent data in real time and transmits it to the data preprocessing module for processing to obtain a standardized dataset; the algorithm processing module extracts features and predicts the third-tier conversion potential of the standardized dataset to generate personalized intervention strategies; the intervention execution module pushes the intervention strategies to the corresponding terminals and monitors the execution of the strategies; the effect feedback module collects intervention-related data, calculates effect evaluation indicators, and feeds them back to the algorithm processing module; the algorithm processing module iteratively optimizes the algorithm based on the feedback data; the data storage module stores all data during the system operation; and the visualization module displays relevant data to support management decisions. Through the operation of this system, the entire process of improving the third-tier conversion rate of insurance agents is automated, intelligent, and precise, effectively improving the third-tier conversion rate and reducing training costs.

[0030] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An algorithm for improving the three-tier conversion rate of insurance agents, characterized in that, Includes the following steps: S1. Data preprocessing: Collect multi-dimensional raw data from insurance agents, clean, complete, normalize, and remove outliers from the raw data to obtain a standardized dataset; the multi-dimensional raw data includes agent basic information data, training data, business development data, customer interaction data, and assessment data. S2. Feature extraction: Based on a standardized dataset, a combination of mutual information and random forest algorithms is used to extract key features that affect the agent's three-way conversion rate and construct a key feature set. The key features include training completion rate, first-order transaction cycle, customer follow-up frequency, customer satisfaction score, assessment pass rate, and the number of referral leads obtained. S3. Three-way potential prediction: Construct an ensemble learning prediction model, input the key feature set into the ensemble learning prediction model, and output the three-way probability and non-transfer risk level of each agent; The ensemble learning prediction model is composed of a weighted fusion of a random forest model, an XGBoost model, and a LightGBM model, and the model weight coefficients are optimized by a grid search algorithm. S4. Personalized intervention strategy generation: Based on the agent's three-way conversion probability, non-conversion risk level, and key characteristic weaknesses, and combined with a preset intervention strategy library, a targeted personalized intervention strategy is generated; the intervention strategy library includes four types of strategies: training, business development, incentive, and supervision. S5. Strategy Iteration and Optimization: Collect real-time behavioral data and three-way transition results data of agents after receiving intervention, calculate the evaluation index of the implementation effect of intervention strategy, and adjust the parameters of integrated learning prediction model and strategy adaptation rules of intervention strategy library according to the evaluation results to achieve dynamic iterative optimization of algorithm.

2. The algorithm for improving the three-tier conversion rate of insurance agents according to claim 1, characterized in that, In step S1, the specific process of data preprocessing is as follows: S11. Data cleaning: Remove duplicate and invalid data from the original data, and use Lagrange interpolation to fill in missing data. S12. Outlier removal: The 3σ principle is used to identify outliers in the original data. Outliers are replaced with the median of the corresponding feature. S13. Data normalization: The min-max normalization method is used to map the cleaned numerical data to the [0,1] interval to eliminate the influence of units. S14. Data encoding: One-hot encoding is used to encode categorical data to obtain a standardized dataset.

3. The algorithm for improving the three-tier conversion rate of insurance agents according to claim 1, characterized in that, In step S2, the specific process of extracting the key features is as follows: S21. Calculate the mutual information value between each feature in the standardized dataset and the three-transformation result using the mutual information method, and select initial features whose mutual information value is greater than a preset threshold. S22. Input the initial features into the random forest algorithm model and calculate the importance score of each initial feature; S23. Sort the features by importance score from high to low, select the top N features as key features, and construct a key feature set; Where N is a preset threshold for the number of features, which is dynamically adjusted according to the agent sample size.

4. The algorithm for improving the three-tier conversion rate of insurance agents according to claim 1, characterized in that, In step S3, the construction process of the ensemble learning prediction model is as follows: S31. Train the random forest model, XGBoost model and LightGBM model respectively, and use the K-fold cross-validation method to optimize the hyperparameters of each individual model. S32. Based on the prediction accuracy of each individual model, the weight coefficients of each model are determined by a grid search algorithm. The weight coefficients are positively correlated with the prediction accuracy of the model. S33. The prediction results of each individual model are weighted and summed with their corresponding weight coefficients to obtain the agent's three-way probability. S34. Based on the probability of three-way conversion, the risk level of non-conversion is divided into high risk (probability of three-way conversion < 30%), medium risk (30% ≤ probability of three-way conversion < 70%) and low risk (probability of three-way conversion ≥ 70%).

5. The algorithm for improving the three-tier conversion rate of insurance agents according to claim 1, characterized in that, In step S4, the specific process of generating the personalized intervention strategy is as follows: S41. For high-risk agents, prioritize matching supervisory strategies and basic training strategies, assign dedicated supervisors to provide one-on-one assistance, supplement basic business development knowledge and sales script training, and set phased small incentive targets. S42. For agents with medium risk, match specialized training strategies with business development support strategies, conduct targeted training based on key characteristics and weaknesses (such as customer follow-up skills and referral methods training), and provide intelligent business development tool support; S43. For low-risk agents, match incentive strategies with advanced training strategies, set up advanced performance incentive awards, conduct advanced professional skills training and team management training, and guide them to actively acquire customers and make referrals. S44. Bind the generated personalized intervention strategy with the agent identifier and output it to the intervention execution module.

6. The algorithm for improving the three-tier conversion rate of insurance agents according to claim 1, characterized in that, In step S5, the evaluation indicators for the effectiveness of the intervention strategy include the post-intervention three-turn rate, strategy execution rate, agent behavior improvement rate, and training cost saving rate; the specific process of algorithm iterative optimization is as follows: S51. Calculate the evaluation indicators for the implementation effect of each intervention strategy, eliminate strategies that do not meet the standards, and optimize the strategy content. S52. Add real-time behavioral data and three-turn result data to a standardized dataset, retrain the ensemble learning prediction model, and adjust the model parameters and feature weights. S53. Update the policy adaptation rules of the intervention policy library to make the policy generation more in line with the actual needs of the agent and realize the closed-loop iteration of the algorithm.

7. A system for improving the three-tier conversion rate of insurance agents, characterized in that, The system is used to implement the insurance agent three-way conversion rate improvement algorithm according to any one of claims 1-6, including: The data acquisition module is used to collect multi-dimensional raw data from insurance agents. The raw data includes basic agent information, training data, sales data, customer interaction data, and performance evaluation data. The data acquisition module interfaces with the insurance company's existing core business system, training system, CRM system, and intelligent sales platform to achieve real-time data collection. The data preprocessing module, connected to the data acquisition module, is used to clean, complete, normalize, remove outliers, and encode the acquired raw data, and output a standardized dataset. The algorithm processing module, connected to the data preprocessing module, includes a feature extraction unit, a prediction model unit, a policy generation unit, and an iterative optimization unit. The feature extraction unit is used to extract key features that affect the agent's three-way conversion rate based on a standardized dataset, and to construct a key feature set. The prediction model unit is used to run an ensemble learning prediction model, inputting a key feature set and outputting the agent's three-way conversion probability and non-conversion risk level; The strategy generation unit is used to generate personalized intervention strategies based on the three-transformation probability, the risk level of non-transformation, and key feature shortcomings, combined with the intervention strategy library. The iterative optimization unit is used to collect real-time data after intervention, evaluate the intervention effect, adjust model parameters and strategy adaptation rules, and realize algorithm iteration. The intervention execution module, connected to the algorithm processing module, is used to receive personalized intervention strategies, push the strategies to agent terminals, supervisor terminals, and relevant management terminals, monitor the strategy execution process, and collect strategy execution data. The intervention execution module supports multiple push methods, including APP message push, SMS reminders, background notifications, and manual supervision assignment. The effect feedback module, connected to the intervention execution module and the algorithm processing module, is used to collect real-time behavioral data, three-way result data and strategy execution feedback of agents after receiving intervention, calculate intervention effect evaluation indicators, and feed the relevant data back to the iterative optimization unit of the algorithm processing module. The data storage module connects to all modules of the system and is used to store standardized datasets, key feature sets, prediction model parameters, intervention strategy library, strategy execution data, effect evaluation data, and three-stage transformation result data. It uses encrypted storage to ensure data security and supports real-time data query and traceability.

8. The insurance agent three-way conversion rate improvement system according to claim 7, characterized in that, The specific data collected by the data acquisition module includes: Agent basic information data: including age, education, work experience, date of employment, team and job level; Training data includes training course completion rate, training assessment results, training participation duration, number of specialized skills training sessions, and training feedback scores; Business development data includes first-order transaction cycle, cumulative number of policies sold, premium amount, number of customer follow-ups, follow-up success rate, number of referral leads obtained, and referral conversion rate; Customer interaction data includes customer inquiry response time, customer satisfaction score, number of customer complaints, policy delivery service quality, and customer follow-up completion rate; The assessment data includes monthly / quarterly performance compliance rate, performance growth rate, completion status of core indicators, and reward and punishment records.

9. The insurance agent three-way conversion rate improvement system according to claim 7, characterized in that, The system also includes a visualization module, which is connected to the algorithm processing module and the data storage module. It is used to display the agent's three-way conversion rate, the distribution of non-conversion risk levels, the implementation status of intervention strategies, and the effect evaluation results in the form of charts. It supports multi-dimensional queries and statistics by team, region, and onboarding cycle, providing decision support for managers.

10. The insurance agent three-way conversion rate improvement system according to claim 7, characterized in that, The system adopts a modular design, supporting seamless integration with the existing information systems of insurance companies without requiring large-scale modifications to the existing systems. It is also scalable, allowing for the addition of data dimensions, optimization of algorithm models, and expansion of intervention strategy types according to the business needs of insurance companies, adapting to the agent management needs of insurance companies of different sizes.

Images