Prediction method and device for insurance product combination linkage
By combining graph neural networks and dilated causal convolution, and updating model parameters using future events, the accuracy problem of predicting the linkage of insurance product portfolios is solved, enabling insurance companies to make scientific decisions and optimize resources.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-15
- Publication Date
- 2026-03-27
AI Technical Summary
Existing technologies cannot accurately predict the interrelationships of insurance product portfolios, leading to unreasonable dynamic premium pricing and increased decision-making risks when simulating demand for new products.
By employing graph neural network-based economic relationship modeling, combined with dilated causal convolution and time trend capture modules, and using future events as monitoring signals, the model parameters are updated through error feedback and dynamic loss calculation to accurately predict the purchase intention of insurance product portfolios.
It provides accurate predictions of the linkage between insurance product portfolios, helping insurance companies to develop scientific product and marketing strategies, optimize resource allocation, improve user satisfaction and market competitiveness, and reduce operating costs.
Smart Images
Figure CN121745368A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of insurance data prediction and analysis, and more particularly to a prediction method and device for the linkage of an insurance product portfolio. BACKGROUND
[0002] In the insurance industry, different insurance products have complex correlations such as mutual replacement, mutual complementation, or mutual exclusion due to the overlap of their liability scopes. For example, term life insurance and whole life insurance in life insurance may have some substitutability, and consumers will choose one of them according to their economic conditions and protection needs; while medical insurance and critical illness insurance in health insurance complement each other, medical insurance can reimburse routine medical expenses, and critical illness insurance provides one-time economic compensation when the insured person suffers from a major illness, and together they provide more comprehensive health protection for the insured person. Such complex correlations make the insurance product portfolio highly dynamic and uncertain.
[0003] With the intensification of competition in the insurance market, there is an urgent need for the industry to accurately predict the linkage of the insurance product portfolio. In dynamic premium pricing, the premium needs to be adjusted in real time according to the user's purchase intention and market changes, but traditional methods cannot provide accurate forward-looking predictions, resulting in unreasonable pricing. When simulating new product demand, the market potential and risks need to be evaluated under conditions such as listing, but existing technologies cannot accurately predict the linkage of new products with existing products and the impact of market environment, increasing the risk of decision-making. SUMMARY
[0004] Based on the above technical problems, the present application provides a prediction method and device for the linkage of an insurance product portfolio to solve the technical problems that existing technologies cannot provide accurate forward-looking predictions for dynamic premium pricing, resulting in unreasonable pricing, and also cannot accurately predict the linkage of new products with existing products and the impact of market environment when simulating new product demand, increasing the risk of decision-making.
[0005] To achieve the above-mentioned purpose, the technical solution adopted by the present application is to provide a prediction method for the linkage of an insurance product portfolio, comprising the following steps: S1 uses an economic relationship modeling module based on a graph neural network to model the economic relationship to form an economic foundation information model; S2 uses a time trend capturing module including an expanded causal convolution to capture multi-scale trend features of the economic foundation information model; S3 injects future events as forward supervision signals into a time graph neural network prediction module, and combines the multi-scale trend features captured in step S2 to update the parameters of the economic foundation information model through an error feedback mechanism and dynamic loss calculation; S4 outputs the prediction results of the user's purchase intention for different insurance products in the future period according to the updated economic foundation information model.
[0006] Further, step S1 comprises: S11 assigns an initial weight to the user based on the regional economic index, and performs secondary weighting on the user weight according to an economic risk label generated from historical claim records; S12 uses cross elasticity between the purchased product and other products having a substitution relationship as the message passing content; S13 dynamically allocates weights in the training process relying on the graph attention mechanism to determine the correlation strength of a certain type of insurance and the regional economic level; Further, the initial weight in step S11 is assigned in the following manner: according to the numerical value of the regional economic index, the initial weight value is allocated to the user according to a preset weight distribution rule.
[0007] Further, the economic risk label in step S11 is generated in the following manner: according to the historical claim records of the user, data such as the number of claims and the amount of claims are counted, and an economic risk label is generated according to a preset risk assessment model.
[0008] Further, the cross elasticity in step S12 is calculated in the following manner: the relationship between price changes and demand changes between the purchased product and other products having a substitution relationship is analyzed, and the cross elasticity value is obtained through a preset cross elasticity calculation formula; The cross elasticity calculation formula is: wherein, represents a product, is the cross elasticity of the product ; represents the percentage change of the demand of the product , and the calculation formula is: wherein, and are the demand of the product before and after the price change, respectively; wherein, represents the percentage change of the price of the product , and the calculation formula is: wherein, and are the price of the product before and after the change, respectively.
[0009] Furthermore, the layered capture method in step S2 is as follows: set up dilated causal convolutional layers with different dilation rates, and perform layer-by-layer convolution operations on the input data to extract trend features at different time scales. The trend features at different time scales include short-term capital flows and economic state evolution.
[0010] Furthermore, the future events mentioned in step S3 include at least one of the following 30 days of economic behavior logs and external economic events.
[0011] Furthermore, the error feedback and dynamic loss calculation in step S3 are performed as follows: S31 compares the predicted results with the actual observed values and calculates the error value; S32 calculates the dynamic loss value based on the preset dynamic loss function; The expression for the dynamic loss function is: For the first The insurance product in time The actual observed values; For the first The insurance product in time The actual predicted value; N is the total number of products; T is the total number of time steps; The dynamic weights for each time step are expressed as follows: For the point in time when a future event occurs; The duration of the event's impact; This is the time-weighted decay coefficient; It is an exponential function; The dynamic weight of the product is expressed as follows: in For products and Cross elasticity, Product association strength coefficient; The dynamic loss function, by introducing dynamic weights for time steps and products, can dynamically adjust the loss value based on the distance to future events and the strength of correlation between products. This allows the model to pay more attention to recent data and products with strong correlations during training, thereby improving the accuracy and adaptability of predictions.
[0012] S33 updates the model parameters based on the dynamic loss value.
[0013] The application provides a prediction device for the combination linkage of insurance products, comprising: An economic relationship modeling unit is configured to construct an economic relationship model based on economic correlation data by means of a graph neural network, determine user weights and regional economic correlation strengths of insurance types, and form an economic foundation information model; A trend feature capturing unit is configured to extract long-term and short-term multi-scale trend features of insurance product purchase intention according to the economic foundation information model by using a structure with hierarchical capturing capability; A dynamic prediction processing unit is configured to introduce future events as supervision signals, update parameters by a loss calculation method with a dynamic weight adjustment mechanism, and adapt the economic foundation information model to market dynamic changes; A result output unit is configured to output prediction results of user purchase intention for different insurance products in a future period of time according to the updated economic foundation information model.
[0014] Compared with the prior art, the prediction method and device for the combination linkage of insurance products have the following beneficial effects: The application provides comprehensive and systematic information support for insurance companies, helps the insurance companies to formulate more scientific and reasonable product strategies, marketing strategies and risk management strategies, avoids decision-making errors caused by one-sided information, and optimizes product combination and sales channels by allocating marketing resources according to the purchase intention prediction of different regions and different products in a future period of time, thereby improving resource utilization efficiency and reducing operating costs. The application uses a graph neural network to deeply analyze economic relationships, comprehensively analyzes multi-source data such as regional economic indexes and user historical claim records, accurately constructs a user economic risk portrait, dynamically quantifies the correlation strength of insurance types and regional economy by means of a graph attention mechanism, and builds a solid economic foundation for prediction; The use of expanded causal convolution can extract trend features of different time scales in layers, grasp the long-term evolution context of insurance product purchase intention, and sensitively capture short-term dynamic changes such as short-term capital flow, so that the prediction result is more in line with market actual fluctuations; The application introduces future events as forward supervision signals into the prediction model, and adjusts model parameters flexibly according to the distance of future events and the correlation strength between products by means of error feedback and a dynamic loss calculation mechanism, so that the model has strong real-time response capability and can capture the influence of newly emerging economic behaviors and external economic events on insurance product purchase intention in time, thereby ensuring that the prediction result is always in resonance with the latest market trends. BRIEF DESCRIPTION OF DRAWINGS
[0015] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.
[0016] Figure 1 The flowchart of the prediction method of the insurance product combination linkage of the present application is shown in the figure. DETAILED DESCRIPTION
[0017] In order to make the technical problems, technical solutions and beneficial effects of the present application more clearly, the following will further describe the present application in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and not to limit the present application.
[0018] It should be noted that when an element is referred to as being "fixed to" or "disposed on" another element, it can be directly on the other element or indirectly on the other element. When an element is referred to as being "connected to" another element, it can be directly connected to the other element or indirectly connected to the other element.
[0019] It should be understood that the terms "length", "width", "upper", "lower", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application.
[0020] In addition, the terms "first", "second" are only for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the technical features indicated. Therefore, the features defined with "first", "second" can explicitly or implicitly include one or more of the features. In the description of the present application, the meaning of "a plurality of", "several" is two or more, unless otherwise specifically limited.
[0021] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present application belongs.
[0022] Please see Figure 1 As shown in the figure, the following will describe the prediction method and device of the insurance product combination linkage provided by the embodiments of the present application. The prediction method of the insurance product combination linkage of the present application comprises the following steps: S1 adopts an economic relationship modeling module based on a graph neural network to model the economic relationship and form an economic foundation information model to provide economic foundation information for subsequent prediction.
[0023] In this embodiment, step S1 includes: S11 assigns an initial weight to the user based on the regional economic index and reweights the user weight according to the economic risk label generated by the historical claim record. Specifically, according to the numerical value of the regional economic index, the user is assigned an initial weight value according to a preset weight distribution rule. For example, the regional economic index can be divided into three intervals of high, medium and low, which correspond to different initial weight coefficients, such as an initial weight of 1.2 for a user in a high economic region, 1.0 for a user in a medium economic region, and 0.8 for a user in a low economic region. According to the historical claim record of the user, the claim frequency, claim amount and other data are counted, and an economic risk label is generated through a preset risk assessment model. For example, a user whose claim frequency exceeds 3 times or whose claim amount exceeds a certain threshold is a high-risk user, and the label is "H"; a user whose claim frequency is between 1-3 times or whose claim amount is moderate is a medium-risk user, and the label is "M"; a user with no claim record or a very low claim amount is a low-risk user, and the label is "L". After obtaining the initial weight and the economic risk label, the system reweights and adjusts the initial weight of the user according to the economic risk label. For example, the weight of a high-risk user is reduced by 20%, the weight of a medium-risk user remains unchanged, and the weight of a low-risk user is increased by 10%.
[0024] S12 uses the cross elasticity between the purchased product and other products with a substitution relationship as the message passing content; specifically, the cross elasticity calculation method is to analyze the relationship between the price change and the demand change between the purchased product and other products with a substitution relationship, and obtain the cross elasticity value through a preset cross elasticity calculation formula, so as to use the calculated cross elasticity value as the message passing content, propagate in the graph neural network, and reflect the substitution relationship between different products.
[0025] The cross elasticity calculation formula is: wherein, represents a product, is the cross elasticity of the product ; represents the percentage change in the demand of the product , and the calculation formula is: wherein, and are the demand quantities of the product before and after the price change, respectively. wherein, denotes the product price change in percentage, calculated as: wherein, and are the product prices before and after the change.
[0026] S13 dynamically allocates weights during the training process to determine the correlation strength of certain insurance products and regional economic levels. In specific implementation, an attention score can be assigned to each node (representing a user or an insurance product), which reflects the importance of the node in the current context. Through weighted summation of the attention scores, the contribution of each node to the final prediction result is obtained.
[0027] S2 adopts a time trend capturing module including dilated causal convolution to capture multi-scale trend features of the economic foundation information model in a hierarchical manner, so as to grasp the long-term and short-term change trends of the insurance product purchase intention. Specifically, the hierarchical capturing manner is to set dilated causal convolution layers with different expansion rates to perform layer-by-layer convolution operations on the input data to extract trend features of different time scales, including short-term capital flow and economic state evolution. For example, three dilated causal convolution layers can be set with expansion rates of 1, 2 and 4 to capture short-term (such as daily), medium-term (such as weekly) and long-term (such as monthly) capital flow and economic state evolution trends.
[0028] S3 injects future events as forward supervision signals into the time graph neural network prediction module, and combines the multi-scale trend features captured in step S2 to update the economic foundation information model parameters through an error feedback mechanism and dynamic loss calculation, so that the model can adapt to the dynamically changing market environment. Specifically, the future events include at least one of the economic behavior logs in the next 30 days and external economic events, which will be injected into the time graph neural network prediction module as forward supervision signals.
[0029] The error feedback and dynamic loss calculation manner is: S31 compares the prediction result with the actual observation value to calculate the error value, which can specifically use indicators such as mean square error (MSE) or mean absolute error (MAE); S32 calculates a dynamic loss value according to a preset dynamic loss function; The expression of the dynamic loss function is: is the insurance product in the time the actual observed value of the product at time is the actual predicted value of the th insurance product at time ; N is the total number of products; T is the total number of time steps; is the time step dynamic weight, and the expression is: is the time point of the future event, is the event impact period; is the time weight decay coefficient; is an exponential function; is the product dynamic weight, and the expression is: wherein is the cross elasticity of the product and , and is the product correlation strength coefficient; The dynamic loss function can dynamically adjust the loss value according to the distance of the future event and the correlation strength between products by introducing the time step dynamic weight and the product dynamic weight, so that the model pays more attention to recent data and products with strong correlation during the training process, thereby improving the accuracy and adaptability of the prediction.
[0030] S33, according to the dynamic loss value, can adopt the gradient descent method, so that the model can adapt to the dynamically changing market environment.
[0031] S4, according to the updated economic foundation information model, outputs the prediction result of the user's purchase intention for different insurance products in a future period of time (such as a future month). The prediction result can be presented in the form of a table, a chart or a report, including the predicted purchase amount, the purchase probability and the possible linkage effect analysis of each insurance product.
[0032] The prediction device for the combination linkage of an insurance product according to the present application comprises: An economic relationship modeling unit is configured to construct an economic relationship model based on economic related data by means of a graph neural network, determine user weights and regional economic correlation strengths of insurance products, and form an economic foundation information model. A trend feature capturing unit is configured to extract long-term and short-term multi-scale trend features of insurance product purchase intention according to the economic foundation information model by using a structure with hierarchical capturing capability. A dynamic prediction processing unit is configured to introduce a future event as a supervision signal, update parameters by using a loss calculation method with a dynamic weight adjustment mechanism, and adapt the economic foundation information model to market dynamic changes. The result output unit outputs a prediction result of the user's purchase intention for different insurance products in a future period of time according to the updated economic foundation information model.
[0033] It can be understood that the parts in the above embodiments can be freely combined or deleted to form different combined embodiments, and the specific content of each combined embodiment will not be described here. After the description, it can be considered that the present application has described each combined embodiment, and can support different combined embodiments.
[0034] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement and improvement within the spirit and principle of the present application shall be included in the protection scope of the present application.
Claims
1. A method for predicting the linkage of insurance product portfolios, characterized in that, Includes the following steps: S1 uses an economic relationship modeling module based on graph neural networks to model economic relationships and form an economic basic information model; S2 employs a time trend capture module, including dilated causal convolution, to capture multi-scale trend features in a hierarchical manner from the economic fundamental information model. S3 uses future events as signals for forward supervision, injects them into the time-map neural network prediction module, and combines them with the multi-scale trend features captured in step S2 to update the parameters of the economic fundamental information model through error feedback mechanism and dynamic loss calculation. Based on the updated economic fundamentals information model, S4 outputs predictions of users' purchasing intentions for different insurance products over a future period.
2. The method for predicting the linkage of insurance product portfolios according to claim 1, characterized in that, Step S1 includes: S11 assigns initial weights to users based on the regional economic index, and then further weights users based on economic risk labels generated from historical claims records. S12 utilizes the cross-elasticity between purchased products and other products with substitutable relationships as the content of the message; S13 uses a graph attention mechanism to dynamically allocate weights during training to determine the correlation strength between a certain type of insurance and the regional economic level.
3. The method for predicting the linkage of insurance product portfolios according to claim 2, characterized in that, The method for assigning the initial weight in step S11 is as follows: based on the magnitude of the regional economic index, the user is assigned an initial weight value according to a preset weight allocation rule.
4. The method for predicting the linkage of insurance product portfolios according to claim 2, characterized in that, The economic risk label in step S11 is generated by: statistically analyzing data such as the number of claims and the amount of claims based on the user's historical claim records, and generating an economic risk label based on a preset risk assessment model.
5. The method for predicting the linkage of insurance product portfolios according to claim 2, characterized in that, The calculation method for the cross elasticity in step S12 is as follows: analyze the relationship between price changes and demand changes between purchased products and other products with substitution relationships, and obtain the cross elasticity value through a preset cross elasticity calculation formula; The formula for calculating cross elasticity is: in, Indicates the product, For relative to the product Cross elasticity; Indicates product The percentage change in demand is calculated using the following formula: in, and These are products Demand before and after price changes; in, Indicates product The percentage change in price is calculated using the following formula: in, and These are products Prices before and after the change.
6. The method for predicting the linkage of insurance product portfolios according to claim 1, characterized in that, The layered capture method in step S2 is as follows: set up dilated causal convolutional layers with different dilation rates, and perform layer-by-layer convolution operations on the input data to extract trend features at different time scales. The trend features at different time scales include short-term capital flows and economic state evolution.
7. The method for predicting the linkage of insurance product portfolios according to claim 1, characterized in that, The future events mentioned in step S3 include at least one of the following 30 days of economic behavior logs and external economic events.
8. The method for predicting the linkage of insurance product portfolios according to claim 1, characterized in that, The error feedback and dynamic loss calculation methods in step S3 are as follows: S31 compares the predicted results with the actual observed values and calculates the error value; S32 calculates the dynamic loss value based on the preset dynamic loss function; The expression for the dynamic loss function is: For the first The insurance product in time The actual observed values; For the first The insurance product in time The actual predicted value; N is the total number of products; T is the total number of time steps; The dynamic weights for each time step are expressed as follows: For the point in time when a future event occurs; The duration of the event's impact; This is the time-weighted decay coefficient; It is an exponential function; The dynamic weight of the product is expressed as follows: in For products and Cross elasticity, Product association strength coefficient; The dynamic loss function, by introducing dynamic weights for time steps and dynamic weights for products, can dynamically adjust the loss value based on the distance to future events and the strength of correlation between products. This allows the model to pay more attention to recent data and products with strong correlations during training, thereby improving the accuracy and adaptability of predictions. S33 updates the model parameters based on the dynamic loss value.
9. A predictive device for the linkage of insurance product portfolios, characterized in that, include: The economic relationship modeling unit is used to construct an economic relationship model based on economic data and with the help of graph neural networks, determine user weights and the strength of the correlation between insurance types and regional economies, and form an economic basic information model. The trend feature capture unit adopts a structure with hierarchical capture capability and extracts long-term and short-term multi-scale trend features of insurance product purchase intentions based on the economic basic information model. The dynamic prediction processing unit introduces future events as monitoring signals and updates parameters through a loss calculation method with a dynamic weight adjustment mechanism, so that the economic basic information model can adapt to dynamic changes in the market. The results output unit outputs the predicted results of users' purchase intentions for different insurance products in the future period based on the updated economic fundamental information model.