Multi-factor fusion insurance product dynamic risk pricing optimization method

Abstract

The invention belongs to the technical field of insurance product pricing, and more specifically relates to a multi-factor fusion insurance product dynamic risk pricing optimization method. Collecting multi-source information data, and carrying out data preprocessing; on the basis of travel trajectory analysis of vehicle-mounted OBD data, in combination with high-risk road sections, driving behaviors and environmental factors, a space-time accumulation risk index is constructed, a graph convolutional network is used for modeling a risk transfer effect in a social circle of vehicle owners, mutual influences among the vehicle owners in the social circle are captured, and a neural network based on multi-task learning is used for establishing a social network model. Dynamic risk, social risk features and other related data are comprehensively processed, the risk probability and loss amount prediction are optimized through a shared network layer, and the insurance pricing is generated in combination with external market data, so that the comprehensiveness and accuracy of the insurance pricing are ensured; the method solves a blind area existing in a conventional pricing method, enables the insurance pricing to be more intelligent and precise, and can reflect the actual risk condition of a vehicle owner in real time.

Description

Technical Field

[0001] This application belongs to the technical field of insurance product pricing, and more specifically, it relates to an optimization method for dynamic risk pricing of insurance products with multi-factor integration. Background Art

[0002] The insurance industry faces the problem of how to accurately assess risks and price reasonably during the pricing process. Traditional insurance pricing methods mainly rely on historical data and statistical models. However, with the explosion of data volume and the continuous development of technology, the pricing of insurance products also urgently needs more intelligent and accurate solutions.

[0003] Traditional insurance pricing methods usually rely on static features, such as basic information like the age, gender, and driving history of the vehicle owner. However, these static features do not fully consider the dynamic risk factors of the vehicle owner, such as real-time driving behavior, vehicle location, and road conditions. This makes the traditional pricing model unable to fully reflect the actual risk situation of the vehicle owner. With the development of in-vehicle technology, the popularization of in-vehicle OBD (On-Board Diagnostic System) devices and real-time meteorological data can provide real-time and detailed driving data and environmental information, providing a new dimension for insurance pricing.

[0004] In addition, existing risk assessment methods often ignore the social connections and risk transmission effects among vehicle owners. Vehicle owners within a social circle may influence each other to a certain extent. For example, the risk of certain driving behaviors may have an extended effect within the social circle, affecting individuals in the entire social network; this factor is an important blind spot for traditional pricing models.

[0005] However, the current conventional pricing method is to price based on the vehicle owner's claim situation and vehicle price. This not only ignores the dynamic risks of the vehicle owner itself but also leads to one-sided pricing methods, and ultimately results in the insurance pricing not being able to match the actual situation. Summary of the Invention

[0006] The present invention provides an optimization method for dynamic risk pricing of insurance products with multi-factor integration, aiming to solve the technical problems of the current one-sided pricing method and the resulting insurance pricing not being able to match the actual situation due to ignoring the dynamic risks of the vehicle owner itself.

[0007] The optimization method for dynamic risk pricing of insurance products with multi-factor integration includes the following steps:

[0008] Step 1: Collect the vehicle owner's policy records, claim data, in-vehicle OBD data, meteorological data, and geographical information data, and preprocess the collected data;

[0009] Step 2: Obtain the owner's travel trajectory based on in-vehicle OBD data. Combine the travel trajectory with high-risk sections to construct the overall section risk of the owner. Then, incorporate the owner's driving behavior and environmental factors to obtain the spatio-temporal cumulative risk index;

[0010] Step 3: Construct a graph structure with the owner as a node and the owner's social circle as the edge between nodes. Use a graph convolutional network to perform convolutional operations on this graph structure, learn the risk conduction effect within the social circle, and generate social risk features;

[0011] Step 4: Based on the preprocessed data, social risk features, and cumulative risk index, construct a neural network for multi-task learning. Learn the common features of the claim probability and loss amount through a shared network layer, and output the claim probability and predicted loss amount;

[0012] Step 5: Generate insurance pricing based on the risk scores and predicted loss amounts output by the neural network for multi-task learning, as well as external market data.

[0013] The present invention collects the owner's insurance policy records, claim data, in-vehicle OBD data, meteorological data, and geographical information data, and performs data preprocessing to provide comprehensive and high-quality data support for subsequent risk assessment. Secondly, through the analysis of travel trajectories based on in-vehicle OBD data, combined with high-risk sections, driving behavior, and environmental factors, a spatio-temporal cumulative risk index is constructed, considering the impact of the owner's dynamic driving behavior and real-time environment, and solving the problem that traditional pricing models fail to fully consider the owner's dynamic risks. In addition, a graph convolutional network is used to model the risk transmission effect in the owner's social circle, effectively capturing the mutual influence among owners within the social circle, and solving the problem that traditional models fail to consider social risk transmission. Finally, based on a neural network for multi-task learning, dynamic risks, social risk features, and other relevant data are comprehensively processed. The prediction of claim probability and loss amount is optimized through a shared network layer, and insurance pricing is generated in combination with external market data, ensuring the comprehensiveness and accuracy of insurance pricing. Therefore, through multi-factor integration, dynamic risk assessment, and social circle modeling, this technical solution solves the blind spots existing in traditional pricing methods, making insurance pricing more intelligent and accurate, and capable of reflecting the actual risk status of the owner in real time.

[0014] Preferably, the step 2 includes the following steps:

[0015] Calibration of high-risk sections: Determine the accident risk value of each section through historical data and real-time traffic monitoring:

[0016]

[0017] where: λ represents the accident attenuation coefficient; N acc(i, t) represents the number of accidents on road segment i at time t; τ represents the current timestamp; k represents the time window size; R acc (i, t) represents the accident risk value on road segment i at time t; e represents the base of the natural logarithm;

[0018] Dynamic risk path allocation: For each trip, calculate the risk weights of the vehicle owner on different road segments:

[0019]

[0020] In the formula: Δt i represents the time for the vehicle owner to drive on this road segment; R path (P) represents the overall path risk of the vehicle owner's driving path;

[0021] Calculation of driving behavior correction factor: Hazardous driving behaviors correct the risk by multiplying the weighted factor with the current risk assessment value:

[0022]

[0023] In the formula: α brake represents the weight factor for hard braking; S brake (i, t) represents the hard braking intensity at time t; S max represents the maximum value of hard braking;

[0024] Weather correction factor:

[0025] f weather (i, t) = β1·f precip (i, t) + β2·f fog (i, t);

[0026] In the formula: represents the precipitation influence factor; represents the haze influence factor; P(i, t) represents the rainfall in area i at time t; V(i, t) represents the visibility in area i at time t; β1 and β2 represent the weight factors;

[0027] Road condition correction factor:

[0028]

[0029] In the formula: Q(i, t) represents the traffic flow in area i at time t; Q max (i) represents the historical maximum flow of area i;

[0030] Calculation of spatio-temporal cumulative risk index:

[0031] R driver (i, t) = R acc (i, t)·(1 + ΔRbrake (i, t) + f weather (i, t) + f trafffc (i, t));

[0032] Where: ΔR brake (i, t) represents the emergency braking correction factor; f weather (i, t) represents the weather correction factor; f traffic (i, t) represents the traffic flow correction factor.

[0033] Preferably, the graph structure is set as G = (V, E, X), where V represents the set of nodes, and each node represents a car owner; E represents the set of edges, and the edges represent the social relationships between car owners, and X represents the feature matrix of the nodes, and each node contains the risk characteristics of the car owner; the node feature matrix X contains the initial risk feature vectors of each car owner; the initial risk feature vectors include the overall path risk R path (P), the emergency braking correction factor ΔR brake (i, t), the weather correction factor f weather (i, t), and the traffic flow correction factor f traffic (i, t); where the adjacency matrix A represents the social relationships between car owners, if there is a social relationship, it is 1, and if not, it is 0.

[0034] Preferably, the graph convolutional network includes a first data processing layer, a second data processing layer, and a global pooling layer;

[0035] The first data processing layer includes a graph convolutional layer and a multi-layer perceptron non-linear transformation layer. Through the graph convolutional layer, a convolution operation is performed on the input feature matrix to convert it into a first feature matrix; the first feature matrix is used as the input of the multi-layer perceptron non-linear transformation layer, and through the multi-layer perceptron non-linear transformation layer, a non-linear transformation is performed on the first feature matrix to obtain the feature matrix output by the first layer;

[0036] The second data processing layer includes a graph convolutional layer and an attention mechanism layer; through the graph convolutional layer, a convolution operation is performed on the feature matrix output by the first layer to obtain a second feature matrix; the attention mechanism layer calculates the attention weights based on the output second feature matrix, and based on the calculated attention weights and the second feature matrix, obtains the feature matrix output by the second layer weighted by the attention mechanism layer;

[0037] Global pooling layer: Summarize the features of all nodes in the graph structure to generate a global representation of the graph and obtain the social risk features.

[0038] Preferably, the loss function of the graph convolutional network is as follows:

[0039]

[0040] Where: λ regularization and λ separation both represent weight parameters;

[0041]

[0042] Where: N represents the number of graphs; R i represents the true social risk feature of the i-th graph; represents the predicted social risk feature of the i-th graph; represents the mean squared error;

[0043]

[0044] Where: represents the global representation of the i-th graph; represents the regularization loss;

[0045]

[0046] Where: (R i -R j ) 2 is used to weight the distance between global representations; represents the global representation of the j-th graph; R j represents the true social risk feature of the j-th graph; ||·||2 represents the Euclidean distance.

[0047] Preferably, the multi-task neural network includes:

[0048] Input layer: The input vector includes social risk features, spatio-temporal cumulative risk indices, and preprocessed data;

[0049] Shared network layer: A multi-layer perceptron is used as the shared network layer, and the input vector is processed through the shared network layer to obtain shared features;

[0050] Probability of occurrence prediction layer: The shared features are processed through the probability of occurrence layer to obtain the probability of occurrence:

[0051]

[0052] Where: W prob represents the weight matrix for probability of occurrence prediction; b prob represents the bias term of the probability of occurrence prediction layer; F shared represents the shared features; σ represents the sigmoid activation function; represents the predicted probability of occurrence;

[0053] Loss amount prediction layer: The shared features are processed through the loss amount prediction layer to obtain the predicted loss amount value:

[0054]

[0055] Wherein: represents the predicted loss amount; W loss represents the weight matrix for loss amount prediction; b loss represents the bias term of the loss amount prediction layer.

[0056] Preferably, the loss function of the multi-task neural network is as follows:

[0057]

[0058] Wherein: λ risk represents the weight parameter of the claim probability loss; represents the loss of the claim probability, and the binary cross-entropy loss function is adopted; represents the loss of the loss amount, and the mean square error loss is adopted; λ loss represents the weight parameter of the loss amount loss.

[0059] Preferably, step 5 includes the following steps:

[0060] Basic pricing: Determine the basic pricing based on the predicted claim probability and the predicted loss amount:

[0061]

[0062] Wherein: represents the predicted claim probability; represents the predicted loss amount; P base represents the basic pricing;

[0063] Basic pricing adjustment: Adjust the basic pricing based on market factors to obtain the adjusted insurance pricing:

[0064] P final = P base ×(1 + α×ΔR + β×C market + γ×I inflation );

[0065] Wherein: ΔR represents the difference from the industry benchmark rate; C market represents the market competition situation; I inflation represents the current inflation rate; α, β, and γ represent adjustment coefficients.

[0066] Preferably, in step 5, the adjustment coefficients α, β, and γ are adjusted by defining a loss function:

[0067]

[0068] Wherein: represents the insurance pricing after the i-th adjustment; represents the actual insurance pricing in history; The optimal adjustment coefficient is found by minimizing the loss function.

[0069] The beneficial effects of the present invention include:

[0070] By collecting various data, the present invention comprehensively reflects the historical behavior of vehicle owners, real-time driving conditions, and external environmental factors, laying a solid foundation for subsequent risk assessment and solving the problem of single and one-sided data sources in traditional pricing models.

[0071] The present invention obtains the travel trajectory of vehicle owners through on-vehicle OBD data, and combines high-risk sections, driving behaviors, and environmental factors to construct a spatio-temporal cumulative risk index, fully considering the dynamic driving behavior of vehicle owners and real-time environmental impacts, and solving the problem that traditional pricing models ignore the dynamic risks of vehicle owners; By combining high-risk sections with specific travel trajectories, the risk status of vehicle owners at different sections and different times can be evaluated more accurately.

[0072] The present invention takes vehicle owners as nodes, introduces the social circle of vehicle owners as the edges between nodes, performs convolutional operations using a graph convolutional network, captures the risk conduction effect within the social circle of vehicle owners, generates social risk features, and based on this, can identify the mutual influence between vehicle owners, especially the transmission effect of risk behaviors within the social circle, solving the problem that traditional models ignore social risk transmission, and thus more comprehensively evaluating the actual risks of vehicle owners.

[0073] The present invention uses a neural network for multi-task learning, learns the common features of the claim probability and loss amount through a shared layer, and respectively outputs the occurrence probability and predicted loss amount, which not only improves the learning efficiency of the model, but also can better capture the potential correlation between the two tasks, solves the sub-optimal problem caused by separation, and improves the accuracy and robustness of risk assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for use in the embodiments or the description of the prior art. Obviously, the following drawings are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0075] Figure 1 It is the overall step block diagram provided by the embodiment of the present invention.

[0076] Figure 2 It is the structural schematic diagram of the graph convolutional network provided by the embodiment of the present invention.

[0077] Figure 3 This is a schematic diagram of the structure of the neural network for multi-task learning provided by an embodiment of the present invention. Detailed implementation manners

[0078] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0079] See Figure 1 As shown, the dynamic risk pricing optimization method for insurance products with multi-factor fusion includes the following steps:

[0080] Step 1: Collect the owner's insurance policy records, claim data, in-vehicle OBD data, meteorological data, and geographical information data, and preprocess the collected data;

[0081] Exemplarily:

[0082] Owner's insurance policy records: Obtain the owner's insurance policy information through the insurance company's system, including policy number, insurance type, historical policy renewal situation, coverage, insurance amount, policy effective and expiration dates, etc.

[0083] Claim data: Obtain the owner's historical claim records from the insurance company, including the occurrence time of each accident, accident type, claim amount, accident liable party, geographical location and road conditions where the accident occurred, etc.

[0084] In-vehicle OBD data: Obtain real-time driving data through the OBD (On-Board Diagnostic) device of the owner's vehicle, including vehicle speed, acceleration, braking situation, engine speed, fuel consumption, driving mileage, driving duration, driving habits (such as hard braking, hard acceleration, turning, etc.), trouble codes, etc.

[0085] Meteorological data: Collect real-time meteorological information related to the owner's travel through a meteorological data interface (such as a meteorological service API), including temperature, humidity, precipitation, wind speed, visibility, etc.

[0086] Geographical information data: Obtain information such as the owner's travel route, geographical coordinates of the road where the owner is located, high-risk sections (such as accident-prone sections, landslide areas, etc.), road grade, traffic flow, etc. through a Geographic Information System (GIS).

[0087] Data cleaning:

[0088] Filter out invalid data: such as filling in missing values and deleting invalid or incomplete records.

[0089] Outlier detection: Identify and correct abnormal data by setting a reasonable range (such as the vehicle speed should not exceed a certain range).

[0090] Data deduplication: Ensure that only one piece of data for the same time period and the same event is retained, and duplicate data is removed.

[0091] Data standardization and normalization:

[0092] Perform standardization processing on all numerical data (such as vehicle speed, driving mileage, meteorological conditions, etc.) so that data with different dimensions fluctuates within the same range, avoiding the influence of certain features on the model performance due to scale differences.

[0093] Use techniques such as Z-score standardization (mean is 0, standard deviation is 1) or Min-Max normalization (data is mapped to the interval [0, 1]) to unify the data distribution.

[0094] Time series processing:

[0095] Since in-vehicle OBD data and meteorological data are usually collected in time series, it is necessary to organize the data according to the specific travel trajectory of the vehicle owner in chronological order to process the driving behavior and meteorological conditions of the vehicle owner at different time periods.

[0096] Handle the missing values in time series data: Fill in the missing time period data through interpolation method to ensure data integrity.

[0097] Feature extraction:

[0098] For in-vehicle OBD data, we extract driving behavior features from it, including average vehicle speed, number of hard brakes, driving distance, and driving duration, etc.;

[0099] For meteorological data, extract the influencing features of the weather from it, such as the influence of rainfall and wind speed on the accident probability;

[0100] Extract features related to road risks from geographic information data, such as the frequency of the vehicle owner driving on high-risk sections, road section types (urban roads, highways, etc.), road condition information, etc.

[0101] Integrate the vehicle owner's insurance policy records, claim data, OBD data, meteorological data, and geographic information data to form a comprehensive data set containing all relevant information.

[0102] Perform unified format conversion on data from different sources (for example, unify the timestamp format, standardize the coordinate system) to ensure data consistency and usability.

[0103] Step 2: Obtain the travel trajectory of the vehicle owner based on in-vehicle OBD data, combine the travel trajectory with high-risk road sections to construct the overall road section risk of the vehicle owner, and then incorporate the driving behavior and environmental factors of the vehicle owner to obtain the spatio-temporal cumulative risk index;

[0104] Step 2 includes the following steps:

[0105] Calibration of high-risk sections: Determine the accident risk value of each section through historical data and real-time traffic monitoring:

[0106]

[0107] where: λ represents the accident attenuation coefficient; N acc (i, t) represents the number of accidents on section i at time t; τ represents the current timestamp; k represents the time window size; R acc (i, t) represents the accident risk value of section i at time t; e represents the base of the natural logarithm;

[0108] Dynamic risk path allocation: For each trip, calculate the risk weights of the vehicle owner on different sections:

[0109]

[0110] where: Δt i represents the time for the vehicle owner to drive on this section; R path (P) represents the overall path risk of the vehicle owner's driving path;

[0111] Calculation of driving behavior correction factor: Hazardous driving behaviors correct the risk by multiplying the weighted factor by the current risk assessment value:

[0112]

[0113] where: α brake represents the weight factor of sudden braking; S brake (i, t) represents the sudden braking intensity at time t; S max represents the maximum value of sudden braking;

[0114] Weather correction factor:

[0115] f weather (i, t) = β1·f precip (i, t) + β2·f fog (i, t);

[0116] where: represents the precipitation influence factor; represents the haze influence factor; P(i, t) represents the rainfall in area i at time t; V(i, t) represents the visibility in area i at time t; β1 and β2 represent weight factors;

[0117] Road condition correction factor:

[0118]

[0119] Where: Q(i, t) represents the traffic flow in area i at time t; Q max (i) represents the historical maximum flow of area i;

[0120] Calculation of the spatio-temporal cumulative risk index:

[0121] R driver (i, t) = R acc (i, t) · (1 + ΔR brake (i, t) + f weather (i, t) + f trafffc (i, t));

[0122] Where: ΔR brake (i, t) represents the hard braking correction factor; f weather (i, t) represents the weather correction factor; f traffic (i, t) represents the traffic flow correction factor.

[0123] In this embodiment, the travel trajectory of the vehicle owner is obtained through in-vehicle OBD data, and combined with high-risk sections, driving behaviors, and environmental factors, a spatio-temporal cumulative risk index is constructed, which fully considers the dynamic driving behaviors of vehicle owners and real-time environmental impacts, and solves the problem that traditional pricing models ignore the dynamic risks of vehicle owners; through the combination of high-risk sections and specific travel trajectories, the risk status of vehicle owners in different sections and at different times can be evaluated more accurately.

[0124] Step 3: Construct a graph structure, where vehicle owners are nodes and the social circles of vehicle owners are the edges between nodes. Use a graph convolutional network to perform convolutional operations on this graph structure to learn the risk conduction effect within the social circle and generate social risk features;

[0125] The graph structure is set as G = (V, E, X), where V represents the set of nodes, and each node represents a vehicle owner; E represents the set of edges, and the edges represent the social relationships between vehicle owners. X represents the feature matrix of nodes, and each node contains the risk features of the vehicle owner; among them, the node feature matrix X contains the initial risk feature vectors of each vehicle owner; the initial risk feature vectors include the overall path risk R path (P), the hard braking correction factor ΔR brake (i, t), the weather correction factor f weather (i, t), and the traffic flow correction factor f traffic (i, t) of vehicle owner i; among them, the adjacency matrix A represents the social relationships between vehicle owners. If there is a social relationship, it is 1, and if not, it is 0.

[0126] In a social circle, a vehicle owner habitually engages in some dangerous driving behaviors, such as speeding, frequent hard braking, frequent lane changes, etc. These behaviors not only increase the personal risk of an accident for the vehicle owner but may also have an impact on other vehicle owners within their social circle. Because in some social circles, vehicle owners influence and imitate each other's behaviors. For example, if a vehicle owner sees that a friend often speeds without being punished, they may inadvertently imitate this behavior, thus increasing their own risk of an accident. Therefore, in this embodiment, by establishing connections through a social network among vehicle owners, the potential risk transmission mechanism among vehicle owners can be captured, enabling a more comprehensive assessment of the indirect risk impact of each vehicle owner within the social circle.

[0127] See Figure 2 As shown, the graph convolutional network includes a first data processing layer, a second data processing layer, and a global pooling layer;

[0128] The first data processing layer includes a graph convolutional layer and a multi-layer perceptron non-linear transformation layer. The input feature matrix is subjected to a convolutional operation through the graph convolutional layer to be converted into a first feature matrix. The first feature matrix is used as the input of the multi-layer perceptron non-linear transformation layer, and the first feature matrix is non-linearly transformed through the multi-layer perceptron non-linear transformation layer to obtain the feature matrix output at the first layer;

[0129] As a possible implementation manner of this embodiment, the multi-layer perceptron includes two fully connected layers. The first fully connected layer receives the output of the graph convolutional layer for non-linear transformation, and the output of the first fully connected layer enters the second fully connected layer. The second fully connected layer further performs non-linear transformation based on the output of the second fully connected layer and then outputs to the second data processing layer.

[0130] In this embodiment, the features output by the graph convolution are non-linearly transformed through a multi-layer perceptron (MLP), enabling the model to capture more complex social network risk patterns; the structure of the multi-layer perceptron can help the model process the features of each node more deeply, thereby improving the prediction accuracy.

[0131] The second data processing layer includes a graph convolutional layer and an attention mechanism layer. The second feature matrix is obtained by performing a convolutional operation on the feature matrix output at the first layer through the graph convolutional layer. The attention mechanism layer calculates attention weights based on the output second feature matrix, and based on the calculated attention weights and the second feature matrix, the feature matrix output at the second layer weighted by the attention mechanism layer is obtained;

[0132] In this implementation, by calculating the attention weights between nodes, the impact of key nodes (such as car owners with higher risks) in the social network on other nodes can be effectively identified and strengthened; the model can automatically focus on car owners with greater influence in the social circle, thereby improving the accuracy of overall risk assessment, especially in the case of a large social circle or complex information.

[0133] Global pooling layer: Aggregate the features of all nodes in the graph structure to generate a global representation of the graph and obtain social risk features. Through the global pooling layer, the node features of the entire graph are aggregated to obtain a global representation; this global representation can comprehensively reflect the risk situation of the entire social circle, enabling the model to understand and evaluate the social risks of car owners from a global perspective.

[0134] The loss function of the graph convolutional network is as follows:

[0135]

[0136] In the formula: λ regularization and λ separation both represent weight parameters;

[0137]

[0138] In the formula: N represents the number of graphs; R i represents the true social risk feature of the i-th graph; represents the predicted social risk feature of the i-th graph; represents the mean squared error;

[0139]

[0140] In the formula: represents the global representation of the i-th graph; represents the regularization loss;

[0141]

[0142] In the formula: (R i -R j ) 2 is used to weight the distance between global representations; represents the global representation of the j-th graph; R j represents the true social risk feature of the j-th graph; ||·||2 represents the Euclidean distance.

[0143] In this embodiment, the loss function includes a regression loss, a regularization loss, and a separation loss. Through the regularization loss, the model can avoid the overfitting problem and ensure the generalization ability of feature learning. The separation loss helps the model better distinguish the risk features of different social circles by weighting the distances between the global representations of nodes. By maximizing the differences between the social risk features of different vehicle owners, the model can optimize the risk assessment of different vehicle owners and improve the risk discrimination among nodes within the social circle.

[0144] Step 4: Construct a multi-task learning neural network based on the preprocessed data, social risk features, and cumulative risk index, and learn the common features of the claim probability and loss amount through a shared network layer, and output the claim probability and the predicted loss amount.

[0145] See Figure 3 As shown, the multi-task neural network includes:

[0146] Input layer: The input vector includes social risk features, spatio-temporal cumulative risk index, and preprocessed data.

[0147] Shared network layer: A multi-layer perceptron is used as the shared network layer, and the input vector is processed through the shared network layer to obtain shared features. The multi-layer perceptron uses three fully connected layers, and the input vector is processed sequentially through the three fully connected layers, and the last fully connected layer outputs the shared features.

[0148] Claim probability prediction layer: Process the shared features through the claim probability layer to obtain the occurrence probability:

[0149]

[0150] Where: W prob Represents the weight matrix for claim probability prediction; b prob Represents the bias term of the claim probability prediction layer; F shared Represents the shared features; σ represents the sigmoid activation function. Represents the predicted claim probability.

[0151] Loss amount prediction layer: Process the shared features through the loss amount prediction layer to obtain the predicted loss amount value:

[0152]

[0153] Where: Represents the predicted loss amount; W loss Represents the weight matrix for loss amount prediction; b loss Represents the bias term of the loss amount prediction layer.

[0154] The loss function of the multi-task neural network is as follows:

[0155]

[0156] Where: λ risk represents the weight parameter of the loss of the risk probability; represents the loss of the risk probability, and the binary cross-entropy loss function is adopted; represents the loss of the loss amount, and the mean square error loss is adopted; λ loss represents the weight parameter of the loss of the loss amount.

[0157] In this embodiment, by designing a multi-task learning framework, the model can share and learn the common features of the risk probability and the loss amount; it can prompt the model to consider both the possibility of a risk occurring and the severity of the loss during the learning process, rather than just dealing with these two tasks separately; through sharing the network layer, the features of the two tasks (risk probability and loss amount) can borrow strength from each other, thereby improving the overall learning effect and avoiding information fragmentation between the two.

[0158] Step 5: Generate an insurance price based on the risk score output by the neural network based on multi-task learning, the predicted loss amount, and external market data.

[0159] The said Step 5 includes the following steps:

[0160] Basic pricing: Determine the basic pricing based on the predicted risk probability and the predicted loss amount:

[0161]

[0162] Where: represents the predicted risk probability; represents the predicted loss amount; P base represents the basic pricing;

[0163] Basic pricing adjustment: Adjust the basic pricing based on market factors to obtain the adjusted insurance price:

[0164] P final =P base ×(1 + α×ΔR + β×C market + γ×I inflation );

[0165] Where: ΔR represents the difference from the industry benchmark rate; C market represents the market competition situation; I inflation represents the current inflation rate; α, β, and γ represent adjustment coefficients;

[0166] The said market competition situation is obtained by weighted summation based on the difference in average pricing from competitors, the difference from the industry benchmark rate, the annual growth rate of the market supply-demand relationship index, and the market penetration rate.

[0167] Exemplary: The difference in average pricing from competitors is calculated as follows:

[0168]

[0169] Where: P avg represents the average pricing of all competitors in the market; represents the premium pricing of the i-th competitor; N represents the number of competitors;

[0170]

[0171] Where: P company represents the company's pricing; ΔP represents the difference in average pricing from competitors.

[0172] Difference from the industry benchmark rate:

[0173]

[0174] Where: R industry represents the industry's standard benchmark rate;

[0175] Supply - demand relationship index of the market:

[0176]

[0177] Where: D market represents the market demand; S market represents the supply;

[0178] Annual growth rate of market penetration:

[0179]

[0180] Where: P penetration，current and P penetration，previous represent the current and previous year's market penetration respectively.

[0181] In step 5, the adjustment coefficients α, β, and γ are adjusted by defining a loss function:

[0182]

[0183] Where: represents the insurance pricing after the i-th adjustment; represents the actual insurance pricing in history; The optimal adjustment coefficients are found by minimizing the loss function.

[0184] In this embodiment, the basic pricing is based on the predicted probability of a claim and the loss amount, enabling the premium for each customer to be highly matched with their actual risk, improving the accuracy and fairness of pricing. Secondly, the dynamic adjustment of market factors (such as industry benchmark rates, market competition conditions, and inflation rates) enables insurance companies to flexibly respond to market fluctuations, adjust premiums in real time in the face of competitive pressures and changes in the economic environment, enhancing market adaptability and competitiveness. By introducing a loss function to optimize the adjustment coefficient, the model can continuously refine the pricing based on historical actual data, minimize the pricing error, and ensure the consistency of pricing with the actual market situation.

[0185] The above are only the preferred embodiments of the present application and are not intended to limit the present application. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present application shall be included within the protection scope of the present application.

Claims

1. A multi-factor integrated insurance product dynamic risk pricing optimization method, characterized in that: The following steps are involved: Step 1: Collect the car owner's insurance policy records, claims data, vehicle OBD data, meteorological data and geographic information data, and pre-process the collected data; Step 2: Obtain the owner's travel trajectory based on the vehicle's OBD data, and construct the owner's overall road section risk based on the travel trajectory and high-risk road sections. Then integrate the owner's driving behavior and environmental factors to obtain the spatiotemporal cumulative risk index. Step 3: Build a graph structure with the car owner as the node and the car owner’s social circle as the edge between the nodes. Use the graph convolutional network to perform convolution operations on the graph structure, learn the risk transmission effect within the social circle, and generate social risk features. Step 4: Construct a multi-task learning neural network based on the preprocessed data, social risk characteristics, and cumulative risk index, learn the common characteristics of the probability of accident and the amount of loss through the shared network layer, and output the probability of accident and the predicted amount of loss; Step 5: Generate insurance pricing based on the risk score and predicted loss amount output by the multi-task learning neural network and external market data.

2. The multi-factor integrated insurance product dynamic risk pricing optimization method according to claim 1 is characterized in that: The step 2 comprises the following steps: High-risk road section calibration: Determine the accident risk value of each road section through historical data and real-time traffic monitoring: Where: λ represents the accident attenuation coefficient; N acc (i, t) represents the number of accidents on road section i at time t; τ represents the current timestamp; k represents the time window size; R acc (i, t) represents the accident risk value of road section i at time t; e represents the base of the natural logarithm; Dynamic risk path allocation: For each trip, the risk weight of the driver on different road sections is calculated: Where: Δt i Indicates the time the owner travels on this road section; R path (P) represents the overall path risk of the vehicle owner’s driving path; Driving behavior correction factor calculation: Risky driving behavior is corrected by multiplying the current risk assessment value with a weighting factor: Where: α brake Indicates the weight factor of emergency braking; S brake (i, t) represents the emergency braking intensity at time t; S max Indicates the maximum value of emergency braking; Weather correction factor: f weather (i,t)=β1·f precip (i,t)+β2·f fog (i,t); Where: represents the precipitation influencing factor; represents the haze impact factor; P(i, t) represents the rainfall in area i at time t; V(i, t) represents the visibility in area i at time t; β1 and β2 represent weight factors; Road Condition Correction Factor: Where: Q(i, t) represents the traffic flow in area i at time t; Q max (i) represents the historical maximum flow of area i; Calculation of time-space cumulative risk index: R driver (i,t)=R acc (i,t)·(1+ΔR brake (i,t)+f weather (i,t)+f traffic (i,t)); Where: ΔR brake (i, t) represents the emergency brake correction factor; f weather (i, t) represents the weather correction factor; f traffic (i, t) represents the traffic flow correction factor.

3. The multi-factor integrated insurance product dynamic risk pricing optimization method according to claim 1 is characterized in that: The graph structure is set as G = (V, E, X), where V represents a node set, each node represents a car owner; E represents an edge set, the edge represents the social relationship between car owners, X represents the feature matrix of the node, and each node contains the risk characteristics of the car owner; the node feature matrix X contains the initial risk feature vector of each car owner; the initial risk feature vector includes the overall path risk R of car owner i path (P), emergency brake correction factor ΔR brake (i, t), weather correction factor f weather (i, t) and traffic flow correction factor f traffic (i, t); where the adjacency matrix A represents the social relationship between car owners, if there is a social relationship, it is 1, and if there is no social relationship, it is 0.

4. The multi-factor integrated insurance product dynamic risk pricing optimization method according to claim 1 is characterized in that: The graph convolutional network includes a first data processing layer, a second data processing layer and a global pooling layer; The first data processing layer includes a graph convolution layer and a multi-layer perceptron nonlinear transformation layer, and performs a convolution operation on the input feature matrix through the graph convolution layer to convert it into a first feature matrix; the first feature matrix is ​​used as the input of the multi-layer perceptron nonlinear transformation layer, and the first feature matrix is ​​nonlinearly transformed through the multi-layer perceptron nonlinear transformation layer to obtain the feature matrix output by the first layer; The second data processing layer includes a graph convolution layer and an attention mechanism layer; the second feature matrix is ​​obtained by performing a convolution operation on the feature matrix output by the first layer through the graph convolution layer; The attention mechanism layer calculates the attention weight based on the output second feature matrix, and obtains the feature matrix of the second layer output weighted by the attention mechanism layer based on the calculated attention weight and the second feature matrix; Global pooling layer: Aggregates the features of all nodes in the graph structure to generate a global representation of the graph and obtain social risk features.

5. The multi-factor integrated insurance product dynamic risk pricing optimization method according to claim 4 is characterized in that: The loss function of the graph convolutional network is as follows: Where: regularization and λ separation Both represent weight parameters; Where: N represents the number of graphs; R i represents the true social risk characteristics of the i-th graph; represents the predicted social risk features of the i-th graph; represents mean square error; Where: represents the global representation of the i-th graph; represents the regularization loss; Where: (R i -R j ) 2 Used to weight the distance between global representations; represents the global representation of the jth graph; R j represents the true social risk feature of the jth graph; ||·||2 represents the Euclidean distance.

6. The multi-factor integrated insurance product dynamic risk pricing optimization method according to claim 1 is characterized in that: The multi-task neural network includes: Input layer: The input vector includes social risk characteristics, spatiotemporal cumulative risk index, and preprocessed data; Shared network layer: A multi-layer perceptron is used as the shared network layer, and the input vector is processed through the shared network layer to obtain shared features; Accident probability prediction layer: The shared features are processed through the accident probability layer to obtain the occurrence probability: Where: W prob b represents the weight matrix of accident probability prediction; prob represents the bias term of the accident probability prediction layer; F shared represents shared features; σ represents the sigmoid activation function; It indicates the predicted probability of accident; Loss amount prediction layer: The loss amount prediction layer processes the shared features to obtain the loss amount prediction value: Where: Indicates the predicted loss amount; W loss b represents the weight matrix of loss amount prediction; b loss Represents the bias term of the loss amount prediction layer.

7. The multi-factor integrated insurance product dynamic risk pricing optimization method according to claim 6 is characterized in that: The loss function of the multi-task neural network is as follows: Where: risk The weight parameter representing the probability of loss; To represent the loss of the probability of accident, the binary cross entropy loss function is used; The loss represents the amount of loss, using mean square error loss; λ loss The weight parameter representing the loss amount.

8. The multi-factor integrated insurance product dynamic risk pricing optimization method according to claim 1 is characterized in that: The step 5 comprises the following steps: Basic pricing: The basic pricing is determined based on the predicted probability of accident and the predicted loss amount: Where: It indicates the predicted probability of accident; represents the predicted loss amount; P base Indicates base pricing; Basic pricing adjustment: Adjust the basic pricing based on market factors to obtain the adjusted insurance pricing: P final =P base ×(1+α×ΔR+β×C market +γ×I inflation ); Where: ΔR represents the difference from the industry benchmark rate; C market Indicates the market competition situation; I inflation represents the current inflation rate; α, β, and γ represent adjustment coefficients.

9. The multi-factor integrated insurance product dynamic risk pricing optimization method according to claim 9 is characterized in that: In step 5, the adjustment coefficients α, β and γ are adjusted by defining a loss function: Where: represents the insurance pricing after the i-th adjustment; Represents actual insurance pricing in history; finds the optimal adjustment coefficient by minimizing the loss function.

Images