An intelligent pricing system for vehicle insurance based on deep learning

Abstract

The application belongs to the technical field of artificial intelligence, and particularly relates to a car insurance intelligent pricing system based on deep learning, which comprises a multi-dimensional data acquisition module, a space-time feature extraction module for extracting features by using a composite neural network, an actuarial distribution mapping module for mapping features into Poisson or gamma distribution parameters, a multi-task optimization loss function module based on maximum likelihood estimation, and a premium automatic generation module; by constructing a space-time feature deep extraction architecture based on a convolutional neural network and a recurrent neural network, the contradiction that a traditional pricing model is difficult to mine nonlinear and time series risk features in high-dimensional driving behavior data can be solved, and this feature extraction mode ensures that the system can capture the real risk portrait of the insured in all directions and multidimensions, which is beneficial to improving the accuracy of risk identification.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically a deep learning-based intelligent pricing system for auto insurance. Background Technology

[0002] With the accelerated digital transformation of the auto insurance industry, actuarial pricing models based on big data and artificial intelligence have become a core research direction in insurance technology. Traditional auto insurance pricing relies on statistical methods such as generalized linear models to deduce payout expectations by analyzing prior features. Although this meets the requirements of actuarial interpretability, it relies on manual feature engineering, makes it difficult to capture the nonlinear coupling relationship of driving behavior, limits the accuracy of risk identification, and cannot achieve differentiated and accurate pricing.

[0003] To compensate for the insufficient representation capabilities of traditional models, existing technologies introduce deep learning technology, which uses its nonlinear fitting capabilities to mine potential risk features in massive trajectory data. However, existing deep learning pricing schemes treat it as a general regression or classification problem, and the black-box feature extraction mechanism of the model is naturally disconnected from the probability distribution constraints of insurance actuarial calculations.

[0004] The loss function of existing deep learning auto insurance pricing technology deviates significantly from the sparsity and long-tail characteristics of insurance claims data. It lacks rigorous probability density function support, making it difficult to convert prediction results into compliant premiums. Furthermore, it underestimates extreme risks, has poor robustness, and insufficient interpretability, making it impossible to pass actuarial compliance audits and thus restricting the scientific and robust nature of auto insurance pricing.

[0005] Therefore, this invention provides a deep learning-based intelligent pricing system for auto insurance. Summary of the Invention

[0006] In order to overcome the shortcomings of the prior art, at least one technical problem raised in the background art is solved.

[0007] The technical solution adopted by this invention to solve its technical problem is: a deep learning-based intelligent pricing system for auto insurance, comprising: The data processing module is used to acquire the operation sequence data and static attribute data of the insured vehicles, and to clean, interpolate and truncate the operation sequence data to convert it into tensor format data. The spatiotemporal feature deep extraction module is used to construct a composite neural network architecture, extract the nonlinear coupling features of the tensor format data in the time and space dimensions, and output a high-dimensional feature vector. The actuarial probability distribution mapping module is used to map the high-dimensional feature vector to a set of parameters of a preset actuarial probability distribution, wherein the preset actuarial probability distribution includes a discrete probability distribution representing the payout frequency and a continuous probability distribution representing the payout intensity. The multi-task joint optimization loss function module is used to construct a composite loss function based on maximum likelihood estimation. The composite loss function includes a payout frequency likelihood term, a payout intensity likelihood term, and a distribution heavy-tail penalty term. The system parameters are updated by minimizing the negative log-likelihood value. The risk premium automatic generation module is used to calculate the expected value of compensation based on the parameter set and generate a suggested premium amount in combination with preset business parameters.

[0008] Preferably, the data processing module includes: The operation sequence data is obtained, including but not limited to high-frequency driving trajectory data, instantaneous speed, acceleration, turning frequency, driving distance, driving time, and driving area; For missing data items in the running sequence data, a sliding window-based interpolation process is performed, using known samples within a preset time window for linear interpolation or spline interpolation. For abnormal deviations in the running sequence data that exceed a preset physical limit threshold, a truncation operation based on the preset threshold is performed; A hierarchical normalization strategy is adopted to perform channel embedding fusion of the processed running sequence data and the static attribute data, transforming it into tensor format data suitable for deep learning network input; The data processing module is responsible for acquiring full operational data of the insured vehicle in real time or near real time, including but not limited to high-frequency driving trajectory data, number of rapid accelerations, number of rapid decelerations, turning frequency, mileage, driving time, and driving area uploaded through the vehicle diagnostic system. At the same time, the data processing module also integrates static dimensional data, including vehicle brand, model, vehicle age, safety configuration and other attribute data, as well as driver's age, gender, driving experience, historical claims records and other prior feature data. In the data preprocessing stage, the system executes standardized cleaning logic, performs sliding window interpolation on missing values, and performs truncation operation based on preset thresholds on abnormal deviation values. Subsequently, the multi-source heterogeneous data is converted into a tensor format suitable for deep learning network input to ensure the numerical stability of subsequent feature extraction.

[0009] Preferably, the composite neural network architecture used in the spatiotemporal feature depth extraction module includes: A one-dimensional convolutional layer is used to perform local feature mining on the tensor format data by sliding convolutional kernels on the time axis to capture driving risk features; A recurrent connection layer, which is connected to the output of the one-dimensional convolutional layer, uses a long short-term memory network or a gated recurrent unit to model the long-term dependency of driving behavior in the time dimension through the internal forget gate and update gate mechanism, and identifies the stability of driving style. The spatiotemporal feature deep extraction module adopts a composite neural network architecture that integrates convolution operators and recurrent connection structures. In this module, for high-frequency driving behavior sequence data, the system performs spatial feature mining on local driving events through one-dimensional convolutional layers and captures sudden risk features in driving behavior by using the sliding operation of convolution kernels. The output of the convolutional layer is fed into a long short-term memory network or a gated recurrent unit. By utilizing its internal forget gate and update gate mechanism, the long-term dependencies of driving behavior in the time dimension are modeled, thereby identifying the stability of the driver's driving style. The output of this module is a high-dimensional feature vector, which highly condenses the complex nonlinear coupling relationship between the insured and their driving environment and driving habits, providing rich semantic support for subsequent actuarial pricing.

[0010] Preferably, the actuarial probability distribution mapping module includes: The compensation cost is broken down into compensation frequency variables and compensation intensity variables; For the payout frequency variable, it is set to follow a Poisson distribution or a negative binomial distribution. The output layer of the actuarial probability distribution mapping module maps the high-dimensional feature vector through an exponential activation function to ensure that the output distribution intensity parameter is positive, which serves as the expected number of claims. For the compensation intensity variable, it is set to follow a gamma distribution or a log-normal distribution. The actuarial probability distribution mapping module outputs the corresponding shape parameters and scale parameters to describe the probability density distribution of the loss amount of a single claim. The actuarial probability distribution mapping module is a key link connecting the deep learning feature space and the actuarial physical space. This module does not directly output risk scores, but maps the high-dimensional feature vectors output by the spatiotemporal feature deep extraction module to a set of parameters for a specific actuarial distribution. The system decomposes compensation costs into two independent random variables: compensation frequency and compensation intensity. For compensation frequency, the system sets it to follow a Poisson distribution or a negative binomial distribution. The output layer of the mapping module uses an exponential activation function to ensure that the output value is always positive, which serves as the strength parameter of the Poisson distribution, i.e., the expected number of claims. For compensation intensity, i.e., the loss amount of a single claim, the system sets it to follow a gamma distribution or a log-normal distribution. The mapping module outputs the corresponding shape and scale parameters. In this way, the output of the deep learning model is given a clear probability density function meaning, ensuring the rigor of the pricing logic.

[0011] Preferably, the composite loss function in the multi-task joint optimization loss function module includes: For the payout frequency likelihood term, calculate the log-likelihood of the observed actual number of claims under the predicted discrete probability distribution of payout frequency; For the payout intensity likelihood term, for the sample in which a payout has occurred, the log-likelihood value of the observed actual payout amount under the predicted continuous probability distribution of payout intensity is calculated; For the heavy-tailed penalty term of the distribution, a weighting coefficient that is positively correlated with the probability density deviation of the high-loss region is introduced into the loss function to improve the system's prediction accuracy for large-amount extreme payout risks; The multi-task joint optimization loss function module guides the parameter updates of the entire pricing system. Unlike traditional mean squared error loss, this system constructs a composite loss function based on maximum likelihood estimation. This loss function consists of a payout frequency likelihood term, a payout intensity likelihood term, and a regularization constraint term. For the payout frequency part, the loss function calculates the log-likelihood of the actual number of claims under the predicted Poisson distribution. For the payout intensity part, it calculates the log-likelihood of the actual payout amount under the predicted gamma distribution only for samples that have paid out. By minimizing the sum of negative log-likelihoods, the system forces the deep learning model to continuously adjust its internal parameters during training, so that the distribution features generated by the model fit the distribution features of the real insurance payout data to the greatest extent. In addition, this module also introduces a penalty term for the heavy-tailed features of the distribution. By enhancing the probability density fitting weight for high-loss regions, it effectively improves the model's prediction accuracy for large-amount extreme payout risks, thereby ensuring the solvency of insurance companies.

[0012] Preferably, the risk premium automatic generation module includes: The average payout frequency and the average payout intensity output by the actuarial probability distribution mapping module are multiplied to obtain the expected value of pure risk loss of the insured vehicle in the future insurance period. The pre-defined expense ratio, risk margin, and profit target are obtained through a closed-loop feedback algorithm. The recommended premium amount is calculated and output based on the expected value of pure risk loss, the expense ratio, the risk margin, and the profit target. The automatic risk premium generation module receives various distribution parameters output by the actuarial probability distribution mapping module. Based on the expected value pricing principle in actuarial science, it executes automated pricing logic. The system multiplies the predicted average payout frequency with the average payout intensity to obtain the expected value of pure risk loss of the insured vehicle during the future insurance period. The automatic risk premium generation module combines preset expense ratios, risk margins, and profit targets, and automatically calculates the final suggested premium amount through a closed-loop feedback algorithm. This amount not only reflects the personalized risk characteristics extracted based on deep learning, but also meets the regulatory requirements for the interpretability and robustness of pricing models because its derivation process is entirely based on rigorous probability distribution assumptions.

[0013] Preferably, the spatiotemporal feature depth extraction module further includes an attention mechanism layer, the attention mechanism layer comprising: Calculate the corresponding weighting coefficients for driving behavior characteristics at different times; Based on the weighting coefficient, the expression of key feature segments that contribute more than a preset threshold to the probability of an accident is automatically identified and enhanced. The key feature segments include speeding behavior segments or violent lane-changing behavior segments in high-speed driving conditions. The weighted feature sequences are aggregated to generate the high-dimensional feature vector. The spatiotemporal feature deep extraction module also includes an attention mechanism layer. This layer automatically identifies the key feature segments that contribute the most to the probability of accidents by calculating the weight coefficients of driving behavior features at different times. For example, sudden braking behavior that occurs frequently in traffic congestion and speeding behavior when driving on highways will be assigned different risk weights through the attention mechanism. This mechanism further enhances the model's ability to distinguish complex driving conditions and ensures the depth and breadth of feature extraction.

[0014] Preferably, the actuarial probability distribution mapping module employs a variational inference mechanism when processing the payout intensity variable, the variational inference mechanism including: While outputting the point estimates of the shape parameter and the scale parameter, the uncertainty range of the parameters is calculated and output. The automatic risk premium generation module dynamically adjusts the risk compensation coefficient based on the level of uncertainty. For risk samples where the uncertainty exceeds a preset deviation threshold, risk hedging is performed by increasing the premium. When processing the payout intensity distribution, the actuarial probability distribution mapping module uses a variational inference mechanism. This mechanism allows the model to output the uncertainty range of the parameters while outputting the point estimate parameters. This enables the automatic risk premium generation module to dynamically adjust the risk compensation coefficient according to the level of prediction uncertainty when calculating the premium. For high uncertainty risks that the model cannot grasp, the premium is automatically increased to achieve risk hedging.

[0015] Preferably, it further includes a model monitoring and online update submodule, wherein the model monitoring and online update submodule includes: During system operation, the degree of deviation between the distribution characteristics predicted by the monitoring model and the actual observed payout distribution is monitored. The quantized value of the offset degree is calculated using a divergence metric algorithm; If the quantized value exceeds the preset judgment threshold, the retraining mechanism is automatically triggered, and the internal weight parameters of the spatiotemporal feature depth extraction module and the actuarial probability distribution mapping module are corrected using incremental compensation data and the latest running sequence data. The model monitoring and online update submodule monitors the deviation between the model's predicted distribution and the actual payout distribution in real time. Once the deviation exceeds the preset judgment threshold, the system will trigger an automatic retraining mechanism to correct the parameters of the spatiotemporal feature depth extraction module and the distribution mapping module using incremental payout data, ensuring that the pricing accuracy remains optimal under changes in the market environment.

[0016] Preferably, the multi-task joint optimization loss function module further includes decoupling constraint terms, which include: Using the mutual information minimization technique, the feature subset used to predict the payout frequency and the feature subset used to predict the payout intensity are decoupled in the latent space of the spatiotemporal feature depth extraction module. The multi-task joint optimization loss function module also includes a decoupling constraint term, which aims to ensure that the payout frequency feature and the payout intensity feature have a certain degree of independence in the latent space, preventing excessive coupling between dependent variables from causing the model to overfit when dealing with rare accident samples. By introducing techniques such as mutual information minimization, the system ensures the logical purity of frequency prediction and intensity prediction, and improves the generalization performance of the model.

[0017] The beneficial effects of this invention are as follows: 1. The intelligent car insurance pricing system based on deep learning described in this invention can solve the contradiction that traditional pricing models are unable to mine nonlinear and temporal risk features in high-dimensional driving behavior data by constructing a spatiotemporal feature deep extraction architecture based on convolutional neural networks and recurrent neural networks. This feature extraction method ensures that the system can capture the true risk profile of the insured in a comprehensive and multi-dimensional way, thereby improving the accuracy of risk identification.

[0018] 2. The intelligent pricing system for auto insurance based on deep learning described in this invention introduces an actuarial probability distribution mapping module to transform the black-box output of the deep learning model into actuarially recognized distribution parameters such as Poisson distribution and gamma distribution. This solves the technical problem of the disconnect between the deep learning model and actuarial logic. This feature ensures that the model output results have clear statistical significance and comply with the actuarial balance principle and industry regulatory standards.

[0019] 3. The intelligent auto insurance pricing system based on deep learning described in this invention solves the problem of insufficient fitting ability of traditional loss functions when dealing with skewed distribution and long-tail features of insurance data by designing a multi-task joint optimization loss function based on maximum likelihood estimation. This loss function, through collaborative modeling of claim frequency and intensity, especially explicit distribution constraints on extreme loss risks, can ensure the system's ability to capture heavy-tail risks and enhance the technical means for insurance companies to prevent solvency risks.

[0020] 4. The intelligent auto insurance pricing system based on deep learning described in this invention eliminates subjective errors caused by manual feature engineering and manual parameter tuning through the automated actuarial mapping logic realized by the risk premium automatic generation module. It can realize the full-link digital flow from raw data collection to final price output. This end-to-end processing mechanism ensures the efficiency and consistency of the pricing process and can support the real-time quotation needs of large-scale concurrency. Attached Figure Description

[0021] The invention will now be further described with reference to the accompanying drawings.

[0022] Figure 1 This is a structural block diagram of a deep learning-based intelligent pricing system for auto insurance in this invention. Figure 2 This is a flowchart of the premium generation process based on actuarial distribution mapping in this invention. Detailed Implementation

[0023] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0024] like Figure 1 and Figure 2 As shown in the embodiment of the present invention, a deep learning-based intelligent pricing system for auto insurance includes: The data processing module is used to acquire the running sequence data and static attribute data of the insured vehicles, and to clean, interpolate and truncate the running sequence data, converting it into tensor format data. The spatiotemporal feature deep extraction module is used to construct a composite neural network architecture, extract the nonlinear coupling features of tensor format data in the time and space dimensions, and output high-dimensional feature vectors. The actuarial probability distribution mapping module is used to map high-dimensional feature vectors to a set of parameters of a preset actuarial probability distribution, wherein the preset actuarial probability distribution includes a discrete probability distribution representing the payout frequency and a continuous probability distribution representing the payout intensity. The multi-task joint optimization loss function module is used to construct a composite loss function based on maximum likelihood estimation. The composite loss function includes a payout frequency likelihood term, a payout intensity likelihood term, and a distribution heavy-tail penalty term. The system parameters are updated by minimizing the negative log-likelihood value. The automatic risk premium generation module is used to calculate the expected value of compensation based on the parameter set and generate the suggested premium amount in combination with the preset business parameters. The progressive relationship between these modules in terms of physical logic and signal flow can ensure a high degree of confidence in the transformation from the original massive heterogeneous data to the final premium pricing result.

[0025] The data processing module is responsible for acquiring all operational data of the insured vehicles in real time or near real time. In this embodiment, the data source is highly multidimensional and dynamic. It mainly collects high-frequency data through the vehicle diagnostic system and the integrated global positioning system sensors. The high-frequency driving trajectory data is collected at a preset sampling frequency and covers key parameters such as latitude and longitude coordinates, instantaneous speed, altitude and satellite positioning quality. The data processing module accurately captures the number of times the vehicle accelerates rapidly, the number of times it decelerates rapidly, and the frequency of turning by taking the derivative of the instantaneous velocity vector in the time dimension. These data do not exist in isolation, but constitute a continuous dynamic driving sequence. Meanwhile, mileage, driving time, and driving area are linked to the corresponding driving segments in real time through a spatiotemporal matching algorithm. In order to build a comprehensive risk profile, the module also integrates static dimension data, including vehicle brand, model, vehicle age, safety configuration and other attribute data, as well as driver's age, gender, driving experience, historical claims records and other prior feature data.

[0026] During the data preprocessing stage, the data processing module executes standardized cleaning logic. Considering the signal loss or drift that may occur from the vehicle sensors in complex environments, the system performs interpolation processing based on a sliding window. Specifically, this system sets a time window of a preset length. When data loss is detected within the window, linear or spline interpolation is performed using known samples within the window to restore the continuity of the driving sequence. For abnormal deviations caused by sensor malfunctions, the system performs a truncation operation based on a preset threshold, that is, data exceeding physical limits is forcibly corrected to a reasonable range. This system adopts a hierarchical normalization strategy to transform multi-source heterogeneous data into a tensor format suitable for deep learning network input. This tensor structure not only preserves the temporal characteristics of the original data, but also performs a preliminary fusion of static vehicle features and dynamic trajectory features through channel embedding technology, ensuring numerical stability in the subsequent feature extraction process.

[0027] The spatiotemporal feature deep extraction module utilizes the hierarchical structure of deep neural networks to automatically extract deep-level risk signals from chaotic driving sequences. In this module, for high-frequency driving behavior sequence data, the system adopts a composite neural network architecture that integrates convolution operators and recurrent connection structures; One-dimensional convolutional layers are applied to the input driving tensor. Their convolutional kernels slide along the time axis. By weighted summation of velocity, acceleration, and turning rate within a local time window, they capture sudden risk features such as emergency avoidance and frequent lane changes. Convolutional layers have translation invariance, which enables the system to effectively identify similar dangerous behaviors that occur at different times.

[0028] Furthermore, in order to address the long-term dependence of driving behavior in the time dimension, the output of the one-dimensional convolutional layer is connected to a long short-term memory network or a gated recurrent unit. Utilizing its internally designed forget gate, input gate, and update gate mechanism, the system can selectively retain information with long-term risk indicators and forget meaningless random disturbances. This recurrent connection structure can model the stability of the driver's driving style, identifying which drivers, although occasionally braking suddenly, maintain a stable overall driving logic, and which drivers have persistent robustness risks. The final output of the spatiotemporal feature deep extraction module is a high-dimensional feature vector. This vector is no longer a collection of original data, but has undergone nonlinear transformation, which highly condenses the complex coupling relationship between the insured and their driving environment and driving habits, providing rich semantic support for subsequent actuarial pricing.

[0029] The spatiotemporal feature deep extraction module also integrates an attention mechanism layer. This layer calculates the weight coefficients of driving behavior features at different times to achieve a refined measurement of risk contribution. In actual driving scenarios, not all trajectory points have equal risk weights. For example, slow braking at low speeds in congested areas usually results in a lower probability of accident damage. Speeding or sudden lane changes that occur while driving on highways often indicate a very high risk of accidents. The attention mechanism layer assigns higher weights to these key feature fragments, enabling the model to automatically focus on truly dangerous driving conditions and improve the depth and breadth of feature extraction.

[0030] The actuarial probability distribution mapping module is the key technical link connecting the deep learning black box model and the physical space of insurance actuarial science in this invention. The core logic of this module is that it no longer allows the deep learning model to directly output a risk score without physical meaning, but forces the model to learn the parameters of a specific actuarial distribution. Specifically, the system decomposes the complex goal of compensation cost into two independent random variables: compensation frequency and compensation intensity, following the classic frequency-intensity modeling paradigm in actuarial science.

[0031] Regarding the frequency of claims, considering that accidents are sparse and discrete events, the system sets them to follow a Poisson distribution or a negative binomial distribution. The output layer of the actuarial probability distribution mapping module maps the input features through an exponential activation function to ensure that the mean parameter of the output is always positive. This is used as the strength parameter of the Poisson distribution, representing the expected number of claims within a unit insurance period. For the payout intensity, i.e. the amount of loss in a single claim, its numerical distribution usually exhibits obvious right skewness and heavy-tailed characteristics. Therefore, the system is set to follow a gamma distribution or a log-normal distribution. The mapping module outputs the corresponding shape parameters and scale parameters. In this way, the high-dimensional vector extracted by the deep learning model is given a clear probability density function meaning, which can not only improve the rigor of the pricing logic in a statistical sense. More importantly, it allows actuaries to directly observe the model's probability distribution predictions for different risk levels, satisfying the requirement that the model's output must have physical meaning.

[0032] The actuarial probability distribution mapping module introduces a variational inference mechanism when processing payout intensity distribution. This mechanism, by learning the posterior distribution of parameters, allows the model to simultaneously output the uncertainty range of parameters while estimating them at the output point. This has significant application value in practical business. When the model encounters new types of vehicles or extreme driving behaviors that it has never seen before, the variational inference layer will output a high uncertainty score. After receiving such a high uncertainty signal, the risk premium automatic generation module can dynamically adjust the risk compensation coefficient according to the preset risk control logic. For high uncertainty risks that the model cannot grasp, it can automatically offset potential prediction errors by increasing the premium.

[0033] The multi-task joint optimization loss function module is responsible for guiding the parameter update logic of the entire pricing system. Its design not only considers the fitting accuracy, but also the special characteristics of insurance business, which is different from the mean squared error loss function commonly used in traditional machine learning. This system constructs a composite loss function based on maximum likelihood estimation. The loss function consists of a payout frequency likelihood term, a payout intensity likelihood term, and a regularization constraint term. For the payout frequency part, the loss function is calculated by observing the log-likelihood of the actual number of claims under the Poisson distribution parameters predicted by the model. Regarding the payout intensity, since only samples that have paid out have loss amounts, the system calculates the log-likelihood of the actual payout amount under the predicted gamma distribution for only the subset of payouts. By minimizing the sum of negative log-likelihoods across the entire dataset, the system forces the deep learning model to continuously adjust its internal convolutional kernel weights and recurrent layer states during training, so that the probability distribution generated by the model matches the distribution of real insurance payout data to the greatest extent in terms of statistical characteristics.

[0034] In addition, to address the common issue of large-scale claims payouts in insurance data, this module has also introduced a penalty term specifically targeting the heavy-tailed characteristics of the distribution. This penalty term enhances the model's sensitivity to extreme payout risks by adding a weighted coefficient related to the probability density deviation of high-loss areas to the loss function. It can solve the problem that traditional models often under-predict when facing extremely low-probability large-loss events, and improve the technical means for insurance companies to prevent solvency risks.

[0035] The multi-task joint optimization loss function module also includes a decoupling constraint term, which uses mutual information minimization technique to ensure that the payout frequency feature and the payout intensity feature have a certain degree of independence in the latent space of the deep neural network. Because the factors affecting the frequency of accidents and the factors affecting the amount of loss from accidents do not completely overlap; By introducing decoupling constraints, we can prevent the model from generating spurious correlations during the learning process and avoid the collapse of the model's overall pricing logic due to fluctuations in a single variable, thereby significantly improving the model's generalization performance and robustness in the face of rare samples.

[0036] The automatic risk premium generation module is the final output of the system. It receives various distribution parameters from the actuarial probability distribution mapping module and executes automated pricing logic. Based on the expected value pricing principle in actuarial science, the system multiplies the predicted average payout frequency with the average payout intensity to obtain the expected pure risk loss of the insured vehicle over a full insurance period in the future. In order to transform pure risk losses into executable commercial premiums, the module also integrates preset expense ratios, risk margins, and the insurance company's annual profit targets through a closed-loop feedback algorithm. The resulting suggested premium amount not only has a high degree of personalization, reflecting the driver's true risk level, but also naturally possesses auditability and compliance due to its rigorous derivation logic following a probability distribution path.

[0037] Furthermore, the deep learning-based intelligent auto insurance pricing system of the present invention also possesses continuous learning and adaptive evolution capabilities. The system integrates a model monitoring and online update submodule, which monitors in real time the deviation between the distribution characteristics predicted by the model and the actual observed claims distribution in the production environment. Once the deviation exceeds a preset judgment threshold, the system will automatically trigger an incremental training mechanism. This mechanism uses the latest collected claims data and driving trajectories to fine-tune the parameters of the spatiotemporal feature deep extraction module and the distribution mapping module. This closed-loop update capability ensures that the pricing system can adapt to changes in the market environment (such as adjustments to traffic regulations and the popularization of autonomous driving technology), so that the pricing accuracy always remains at an industry-leading level.

[0038] Example: In this embodiment, the system of the present invention is applied to the data of 1 million insured vehicles of a large property insurance company. The input data includes the high-frequency OBD trajectory of each vehicle over the past year, vehicle static attributes, and basic driver information. The system configuration parameters are as follows: The spatiotemporal feature depth extraction module adopts 3 layers of one-dimensional convolution followed by 2 layers of bidirectional LSTM; The actuarial probability distribution mapping module sets the payout frequency to a negative binomial distribution and the payout intensity to a gamma distribution. The model is trained using the Adam optimizer with an initial learning rate of 0.001, and is combined with a cosine annealing strategy.

[0039] Comparative example: The comparison model adopts the generalized linear model commonly used in the insurance industry, which uses the same dataset; However, manual feature engineering is required to extract aggregated features such as annual mileage, peak-hour driving ratio, and number of emergency brakings. The GLM model also adopts a frequency-intensity two-stage modeling approach. The frequency is represented by a Poisson distribution, and the intensity is represented by a gamma distribution.

[0040] Table 1: Comparison between Examples and Comparative Examples:

[0041] As shown in Table 1, the deep learning-based intelligent pricing system for auto insurance provided by this invention significantly outperforms traditional solutions in several core dimensions. Regarding the accuracy of risk identification, thanks to the spatiotemporal feature deep extraction module's mining of nonlinear features, this invention improves the Gini coefficient for payout frequency prediction by 72.7%, meaning the system can more accurately identify high-risk drivers. By introducing a multi-task joint optimization loss function and a heavy-tail penalty term, the system's capture rate of extreme risks is improved by 127.2%, which is crucial for insurance companies to control underwriting quality and avoid devastating payouts. Due to the adoption of an actuarial probability distribution mapping module, this invention maintains the high accuracy of deep learning while ensuring its output has rigorous statistical meaning, guaranteeing the transparency and compliance of pricing logic at the regulatory level, and successfully resolving the contradiction between the black-box nature of deep learning and actuarial compliance.

[0042] In terms of specific physical implementation, the system of this invention can be deployed on a high-performance cloud computing platform. The data processing module receives data streams from tens of millions of vehicle terminals through a distributed message queue and uses a streaming processing framework to complete real-time cleaning and tensor transformation. The spatiotemporal feature deep extraction module and the subsequent actuarial mapping module are deployed on a GPU computing cluster and use the operator library provided by the deep learning framework to achieve parallel acceleration of computing. The risk premium automatic generation module connects with the core business system of the insurance company through a standard RESTful API interface to achieve a real-time quotation response in seconds.

[0043] In summary, this invention, by embedding actuarial distributed logic within a deep learning architecture, ensures the model's powerful nonlinear representation capabilities while also maintaining the rigor of actuarial science and regulatory compliance. Through rigorous logical mapping and signal feedback, the various modules of the system form a closed-loop technology capable of autonomous evolution and precise pricing. This deep learning pricing paradigm based on probability distribution mapping provides a complete technical solution for achieving differentiated and precise auto insurance pricing.

[0044] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A deep learning-based intelligent pricing system for auto insurance, characterized in that, include: The data processing module is used to acquire the operation sequence data and static attribute data of the insured vehicles, and to clean, interpolate and truncate the operation sequence data to convert it into tensor format data. The spatiotemporal feature deep extraction module is used to construct a composite neural network architecture, extract the nonlinear coupling features of the tensor format data in the time and space dimensions, and output a high-dimensional feature vector. The actuarial probability distribution mapping module is used to map the high-dimensional feature vector to a set of parameters of a preset actuarial probability distribution, wherein the preset actuarial probability distribution includes a discrete probability distribution representing the payout frequency and a continuous probability distribution representing the payout intensity. The multi-task joint optimization loss function module is used to construct a composite loss function based on maximum likelihood estimation. The composite loss function includes a payout frequency likelihood term, a payout intensity likelihood term, and a distribution heavy-tail penalty term. The system parameters are updated by minimizing the negative log-likelihood value. The risk premium automatic generation module is used to calculate the expected value of compensation based on the parameter set and generate a suggested premium amount in combination with preset business parameters.

2. The intelligent auto insurance pricing system based on deep learning according to claim 1, characterized in that, The data processing module includes: The operation sequence data is acquired, including but not limited to high-frequency driving trajectory data, instantaneous speed, acceleration, turning frequency, driving distance, driving time, and driving area; For missing data items in the running sequence data, a sliding window-based interpolation process is performed, using known samples within a preset time window for linear interpolation or spline interpolation. For abnormal deviations in the running sequence data that exceed a preset physical limit threshold, a truncation operation based on the preset threshold is performed; A hierarchical normalization strategy is adopted to perform channel embedding fusion between the processed running sequence data and the static attribute data, transforming it into tensor format data suitable for deep learning network input.

3. The intelligent auto insurance pricing system based on deep learning according to claim 1, characterized in that, The composite neural network architecture used in the spatiotemporal feature depth extraction module includes: A one-dimensional convolutional layer is used to perform local feature mining on the tensor format data by sliding convolutional kernels on the time axis to capture driving risk features; The recurrent connection layer, which is connected to the output of the one-dimensional convolutional layer, uses a long short-term memory network or a gated recurrent unit to model the long-term dependencies of driving behavior in the time dimension through the internal forget gate and update gate mechanism, and identifies the stability of driving style.

4. The intelligent auto insurance pricing system based on deep learning according to claim 3, characterized in that, The spatiotemporal feature depth extraction module further includes an attention mechanism layer, which includes: Calculate the corresponding weighting coefficients for driving behavior characteristics at different times; Based on the weighting coefficient, the expression of key feature segments that contribute more than a preset threshold to the probability of an accident is automatically identified and enhanced. The key feature segments include speeding behavior segments or violent lane-changing behavior segments in high-speed driving conditions. The weighted feature sequences are aggregated to generate the high-dimensional feature vector.

5. The intelligent auto insurance pricing system based on deep learning according to claim 1, characterized in that, The actuarial probability distribution mapping module includes: The compensation cost is broken down into compensation frequency variables and compensation intensity variables; For the payout frequency variable, it is set to follow a Poisson distribution or a negative binomial distribution. The output layer of the actuarial probability distribution mapping module maps the high-dimensional feature vector through an exponential activation function to ensure that the output distribution intensity parameter is positive, which serves as the expected number of claims. For the compensation intensity variable, it is set to follow a gamma distribution or a log-normal distribution. The actuarial probability distribution mapping module outputs the corresponding shape parameters and scale parameters to describe the probability density distribution of the loss amount of a single claim.

6. The intelligent auto insurance pricing system based on deep learning according to claim 5, characterized in that, When processing the payout intensity variable, the actuarial probability distribution mapping module employs a variational inference mechanism, which includes: While outputting the point estimates of the shape parameter and the scale parameter, the uncertainty range of the parameters is calculated and output. The automatic risk premium generation module dynamically adjusts the risk compensation coefficient based on the level of uncertainty. For risk samples where the uncertainty level exceeds a preset deviation threshold, risk hedging is performed by increasing the premium.

7. The intelligent auto insurance pricing system based on deep learning according to claim 1, characterized in that, The composite loss function in the multi-task joint optimization loss function module includes: For the payout frequency likelihood term, calculate the log-likelihood of the observed actual number of claims under the predicted discrete probability distribution of payout frequency; For the payout intensity likelihood term, for the sample in which a payout has occurred, the log-likelihood value of the observed actual payout amount under the predicted continuous probability distribution of payout intensity is calculated; For the heavy-tailed penalty term of the distribution, a weighted coefficient that is positively correlated with the probability density deviation of the high-loss region is introduced into the loss function to improve the system's prediction accuracy for large-amount extreme payout risks.

8. The intelligent auto insurance pricing system based on deep learning according to claim 7, characterized in that, The multi-task joint optimization loss function module further includes decoupling constraint terms, which include: By utilizing the mutual information minimization technique, the feature subset used to predict the payout frequency and the feature subset used to predict the payout intensity are decoupled within the latent space of the spatiotemporal feature depth extraction module.

9. The intelligent auto insurance pricing system based on deep learning according to claim 1, characterized in that, The automatic risk premium generation module includes: The average payout frequency and the average payout intensity output by the actuarial probability distribution mapping module are multiplied to obtain the expected value of pure risk loss of the insured vehicle in the future insurance period. The pre-defined expense ratio, risk margin, and profit target are obtained through a closed-loop feedback algorithm. The recommended premium amount is calculated and output based on the expected value of pure risk loss, the expense ratio, the risk margin, and the profit target.

10. The intelligent auto insurance pricing system based on deep learning according to claim 1, characterized in that, It also includes a model monitoring and online update submodule, which includes: During system operation, the degree of deviation between the distribution characteristics predicted by the monitoring model and the actual observed payout distribution is monitored. The quantized value of the offset degree is calculated using a divergence metric algorithm; If the quantized value exceeds the preset judgment threshold, the retraining mechanism is automatically triggered, and the internal weight parameters of the spatiotemporal feature depth extraction module and the actuarial probability distribution mapping module are corrected using incremental compensation data and the latest running sequence data.

Images