Loss prediction method, apparatus, medium, and device

Abstract

The application relates to the technical field of artificial intelligence, and is applied to the financial field, and discloses a loss prediction method, device, equipment and medium, which comprises the following steps: acquiring multi-source risk data corresponding to a to-be-predicted insurance policy, performing feature engineering processing on the multi-source risk data to generate an input feature set; inputting the input feature set into a trained claim frequency prediction model and a trained claim intensity prediction model respectively, and outputting an expected claim frequency and an expected claim intensity; performing distribution verification on the expected claim frequency and the expected claim intensity to obtain a verification result; performing bias calibration on the expected claim frequency and the expected claim intensity according to the verification result to obtain calibrated frequency distribution parameters and intensity distribution parameters; constructing a compound loss distribution based on the frequency distribution parameters and the intensity distribution parameters, calculating a loss distribution expectation value; and determining a risk loss according to the loss distribution expectation value and the risk exposure of the to-be-predicted insurance policy, so that the prediction accuracy and reliability are improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology and is applied to financial technology scenarios, particularly to a loss prediction method, apparatus, device, and medium. Background Technology

[0002] The business characteristics of employer's liability insurance determine that there are significant differences in the industries, sectors, personnel characteristics, and related risk points involved. Currently, the industry mostly uses the traditional experience rate method and the generalized linear model (GLM) to predict pure risk premiums. By analyzing and calculating the loss ratio according to the dimensions of industry type, number of employees, average monthly salary, accident rate, regional differences, job category, region, and company size, it is required to have relatively complete historical loss data and stable risk characteristics of the company or industry.

[0003] In related technologies, the traditional experience-based premium rate method uses the number of employees or total wages as risk exposure indicators, such as the average number of employees or total labor costs per year. Different industries, regions, and positions have significantly different risk levels, requiring further segmentation (manufacturing and construction, etc.). This necessitates collecting historical claims records, including the compensation amount for each accident (medical expenses, disability benefits, and lost wages, etc.), the number and type of accidents (work-related injuries, occupational diseases, and third-party liability, etc.), case processing cycles, and payout rates. This allows for the classification of various risks, and the average payout cost for each type of risk is calculated based on the total payout amount and total risk exposure, thus yielding the experience premium rate for each risk. In contrast, the Generalized Linear Model (GLM) captures the nonlinear relationships and interactions between multidimensional risk factors, modeling complex payout behavior as interpretable statistical relationships. This can improve the scientific nature of pricing, risk differentiation, and underwriting profitability.

[0004] However, both existing approaches fail to adequately consider the impact of various relatively new risk factors on pure risk premium forecasting. They are sensitive to small sample sizes or highly volatile data, lack foresight, have limited model generalization ability, suffer from large prediction errors, and are prone to overfitting. Furthermore, they perform poorly in applications involving new businesses, new products, and new scenarios, thus affecting the overall accuracy and reliability of forecasting. Therefore, there is an urgent need for a loss forecasting method that can overcome these shortcomings and improve the accuracy and reliability of loss forecasting. Summary of the Invention

[0005] This invention provides a loss prediction method, apparatus, computer equipment, and medium, aiming to solve the technical problems in current technology, such as limited model generalization ability, large prediction error, easy overfitting, poor application effect for new businesses, new products, and new scenarios, which affect the overall prediction accuracy and reliability, and improve the prediction accuracy and reliability of loss prediction.

[0006] Firstly, a loss prediction method is provided, including: Obtain multi-source risk data corresponding to the policy to be predicted, perform feature engineering processing on the multi-source risk data, and generate an input feature set; The input feature set is input into the trained payout frequency prediction model and the trained payout intensity prediction model respectively to predict payout frequency and payout intensity, and output the expected payout frequency and expected payout intensity respectively. The expected payout frequency and the expected payout intensity are subjected to distribution verification to obtain the verification results; Based on the verification results, deviation calibration is performed on the expected payout frequency and the expected payout intensity to obtain the calibrated frequency distribution parameters and intensity distribution parameters; A composite loss distribution is constructed based on the frequency distribution parameters and the intensity distribution parameters, and the expected value of the loss distribution is calculated. The risk loss is determined based on the expected value of the loss distribution and the risk exposure corresponding to the policy to be predicted.

[0007] Secondly, a loss prediction device is provided, comprising: The feature processing unit is used to acquire multi-source risk data corresponding to the policy to be predicted, perform feature engineering processing on the multi-source risk data, and generate an input feature set. The input unit is used to input the input feature set into the trained payout frequency prediction model and the trained payout intensity prediction model respectively to predict the payout frequency and payout intensity, and output the expected payout frequency and expected payout intensity respectively. A verification unit is used to perform distribution verification on the expected payout frequency and the expected payout intensity to obtain verification results. The calibration unit is used to perform deviation calibration on the expected payout frequency and the expected payout intensity according to the verification result, so as to obtain the calibrated frequency distribution parameters and intensity distribution parameters. The calculation unit is used to construct a composite loss distribution based on the frequency distribution parameters and the intensity distribution parameters, and to calculate the expected value of the loss distribution; The determining unit is used to determine the risk loss based on the expected value of the loss distribution and the risk exposure corresponding to the policy to be predicted.

[0008] Thirdly, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the loss prediction method described above.

[0009] Fourthly, a computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, implements the steps of the above-described loss prediction method.

[0010] In the above-described loss prediction method, apparatus, computer equipment, and storage medium, the following scheme can be implemented: First, multi-source risk data corresponding to the policy to be predicted is acquired. Then, feature engineering processing is performed on the multi-source risk data to generate an input feature set. This input feature set is then input into a trained claim frequency prediction model and a trained claim intensity prediction model to predict the claim frequency and claim intensity, respectively, outputting the expected claim frequency and expected claim intensity. Distribution verification is performed on the expected claim frequency and expected claim intensity to obtain verification results. Based on the verification results, deviation calibration is performed on the expected claim frequency and expected claim intensity to obtain calibrated frequency distribution parameters and intensity distribution parameters. A composite loss distribution is constructed based on the frequency distribution parameters and intensity distribution parameters, and the expected loss distribution value is calculated. Finally, based on the expected loss distribution value and the risk exposure corresponding to the policy to be predicted, the risk loss is determined. This invention acquires multi-source risk data of the policy to be predicted and performs feature engineering processing to generate an input feature set. Feature engineering breaks through the limitations of traditional solutions that rely solely on a few basic risk indicators, enabling the integration and standardized transformation of multi-dimensional emerging risk factors. This addresses the shortcomings of incomplete risk characterization at the data input level. Secondly, a parallel prediction architecture using dual-branch deep learning models for payout frequency and payout intensity is employed. Leveraging the non-linear feature mining capabilities of deep learning, it learns the implicit correlation between risk factors and losses. Simultaneously, it models data with different distribution characteristics for accident frequency and payout amount, enhancing the model's adaptability to volatile data and small-sample scenarios, and improving generalization performance in new scenarios. Finally, a closed-loop constraint mechanism constructed through distribution verification, bias calibration, and composite loss distribution is used to verify and correct the deep learning output results using the inherent statistical laws of insurance losses. This suppresses the overfitting risk of purely data-driven models and ensures that the prediction results conform to the statistical rationality of actuarial science. Thus, by addressing the shortcomings of existing technologies at three levels—data dimension, modeling capability, and result constraints—a comprehensive improvement in the accuracy and reliability of loss prediction is achieved. Attached Figure Description

[0011] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a schematic diagram of an application environment for the loss prediction method in one embodiment of the present invention; Figure 2 This is a flowchart illustrating a loss prediction method in one embodiment of the present invention; Figure 3 yes Figure 2 A schematic diagram of a specific implementation method for step S10; Figure 4 yes Figure 3 A flowchart illustrating a specific implementation of step S11; Figure 5 yes Figure 2 A schematic diagram of a specific implementation method for step S20; Figure 6 yes Figure 2 A schematic diagram of a specific implementation method for step S30; Figure 7 yes Figure 2 A schematic diagram of a specific implementation method for step S60; Figure 8 This is a schematic diagram of a loss prediction device in one embodiment of the present invention; Figure 9 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention; Figure 10 This is another structural schematic diagram of a computer device according to one embodiment of the present invention. Detailed Implementation

[0013] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] The loss prediction method provided in this invention can be applied to, for example, Figure 1In this application environment, the client communicates with the server via a network. The client can upload the policy to be predicted to the server; the server obtains multi-source risk data corresponding to the policy to be predicted, performs feature engineering on the multi-source risk data, and generates an input feature set; the input feature set is then input into a trained claim frequency prediction model and a trained claim intensity prediction model to predict claim frequency and claim intensity, respectively, and outputs the expected claim frequency and expected claim intensity; the expected claim frequency and expected claim intensity are validated to obtain validation results; based on the validation results, the expected claim frequency and expected claim intensity are calibrated to obtain calibrated frequency distribution parameters and intensity distribution parameters; a composite loss distribution is constructed based on the frequency distribution parameters and intensity distribution parameters, and the expected loss distribution value is calculated; based on the expected loss distribution value and the risk exposure corresponding to the policy to be predicted, the risk loss is determined. This invention, by obtaining multi-source risk data of the policy to be predicted and performing feature engineering, highlights... This approach breaks away from the limitations of traditional solutions that rely solely on a few basic risk indicators, integrating and standardizing multi-dimensional emerging risk factors to address the shortcomings of incomplete risk characterization at the data input level. Secondly, it employs a parallel prediction architecture using dual-branch deep learning models for payout frequency and payout intensity. Leveraging the non-linear feature mining capabilities of deep learning, it learns the implicit correlation between risk factors and losses. Simultaneously, it models data with different distribution characteristics for accident frequency and payout amount, enhancing the model's adaptability to volatile data and small-sample scenarios, and improving generalization performance in new scenarios. Finally, through a closed-loop constraint mechanism constructed using distribution verification, bias calibration, and composite loss distribution, it verifies and corrects the deep learning output results with the inherent statistical laws of insurance losses, suppressing the overfitting risk of purely data-driven models and ensuring that the prediction results conform to the statistical rationality of insurance actuarial science. Thus, it addresses the shortcomings of existing technologies layer by layer from three levels: data dimension, modeling capability, and result constraints, achieving an overall improvement in the accuracy and reliability of loss prediction.

[0015] Please see Figure 2 As shown, Figure 2 A flowchart illustrating the loss prediction method provided in this embodiment of the invention includes the following steps: S10: Obtain multi-source risk data corresponding to the policy to be predicted, perform feature engineering on the multi-source risk data, and generate input feature set.

[0016] Among them, the policy to be predicted refers to the liability insurance policy that requires risk loss calculation, which is the object of this method and includes basic underwriting information such as the insurer, coverage amount, and coverage period; multi-source risk data refers to various data sets related to the risk level of the policy collected from multiple data channels, covering multiple dimensions such as the policy's own attributes, insurer characteristics, and external risk environment, and is the data foundation for risk calculation; feature engineering refers to the data processing process of deriving, combining, transforming, and standardizing the original multi-source risk data, which transforms the scattered and heterogeneous original data into structured features that can be recognized by the model and can effectively represent the risk pattern; input feature set refers to the standardized feature data set formed after feature engineering, which has a unified data format and dimensions, and serves as the unified input for the subsequent prediction model.

[0017] Specifically, firstly, all risk-related data associated with the policy to be predicted is retrieved from multiple data sources such as business systems and external databases through data interfaces to form a multi-source risk data set; then, feature engineering operations such as feature derivation, feature combination, and numerical standardization are performed on the multi-source risk data in sequence to transform the unstructured and multi-caliber raw data into feature data in a unified format that can represent the risk pattern, and finally, the data is spliced ​​together to form a structured input feature set, which serves as the unified input for subsequent model inference.

[0018] In some embodiments, such as Figure 3 As shown, multi-source risk data includes at least policy core data, personnel exposure data, and external supplementary data. Step S10, which involves performing feature engineering on the multi-source risk data to generate an input feature set, specifically includes the following steps S11 to S14: S11: Extract risk characteristics and time characteristics from core policy data, personnel exposure data, and external supplementary data.

[0019] Among them, core policy data refers to business data stored in the underwriting business system that records the basic underwriting elements of the policy to be predicted. It is the core data source reflecting the underwriting conditions and basic risk attributes of the policy. Personnel exposure data refers to relevant data recording the job attributes, personnel structure and risk exposure of the insured personnel of the insured enterprise. It is the core data source for characterizing the risk level of personnel. External supplementary data refers to external environmental data that comes from outside the policy business system and is related to the risk level of the policy. It is used to supplement external risk dimensions such as industry, region and macro environment that cannot be covered by the policy's endogenous data. Risk characteristics refer to numerical characteristics extracted and derived from multi-source risk data. They can directly quantify the level of risk of the policy and are the core basic characteristics for risk prediction. Time characteristics refer to periodic characteristics extracted from the time attribute fields of the data. They are used to capture the periodic fluctuations of risk with seasonal and time periods.

[0020] Specifically, the acquired multi-source risk data is divided into three categories based on data source and business attributes. Core policy data is extracted from policy information in the underwriting business system, personnel exposure data is extracted from the list of insured personnel and job information of the insured company, and external supplementary data is obtained synchronously from third-party industry databases and public statistical platforms. These three types of data together constitute the complete raw data foundation for feature engineering processing. Based on the raw fields of core policy data, personnel exposure data, and external supplementary data, quantifiable risk features are generated through business rule calculation and attribute extraction. At the same time, periodic attributes such as months and quarters are extracted from the time fields in the data to generate time features that can reflect the time dimension of risk fluctuation patterns, thus completing the derivation and extraction of basic features.

[0021] S12: Perform feature cross-processing on the target industry of the core policy data and the target occupation of the personnel exposure data to obtain combined features.

[0022] Among them, the target industry refers to the industry classification dimension of the insured enterprise corresponding to the policy to be predicted, which is the core classification attribute affecting the risk level of liability insurance; the target occupation refers to the occupational classification dimension of the employees covered by the policy to be predicted, which is the core classification attribute determining the risk level of the occupation; feature cross-processing refers to the data processing operation of combining the values ​​of two independent categorical features to generate new higher-order categorical features, which is used to explore the joint risk patterns after the superposition of multiple features; combined features refer to the higher-order features generated after feature cross-processing, which can characterize the joint risk attributes after the combination of different risk factors and reflect the interaction effect of risk factors.

[0023] Specifically, the industry classification field in the core policy data is extracted as the target industry, and the occupation classification field in the personnel exposure data is extracted as the target occupation. The values ​​of the two types of features are combined one by one to generate a high-order category feature, i.e., a combined feature, of the "industry-occupation" joint dimension. The joint risk pattern of industry and occupation is made explicit into a feature dimension that the model can recognize.

[0024] S13: Combine risk features, time features, and combined features to obtain a combined feature set.

[0025] Among them, the spliced ​​feature set refers to the full set of features formed by aligning risk features, time features, and combination features dimensionally and splicing them horizontally according to the same policy subject. It is the input object for dimensionality reduction processing.

[0026] Specifically, taking the policy to be predicted as the sole subject, the aforementioned risk features, time features, and combination features are aligned in dimensions and stitched together horizontally to ensure that all feature dimensions of the same policy correspond one-to-one and are integrated to form a stitched feature set with a unified structure.

[0027] S14: Perform dimensionality reduction on the concatenated feature set to obtain the input feature set.

[0028] Dimensionality reduction refers to the data processing process of compressing the dimensions of a high-dimensional feature set while preserving core risk information as much as possible. This is used to reduce feature redundancy and lower the computational complexity of subsequent models.

[0029] Specifically, a pre-trained unsupervised dimensionality reduction model is used to compress sparse combined features with high dimensionality in the spliced ​​feature set. While retaining core risk information, the total dimensionality of the features is reduced and feature redundancy is eliminated. Finally, an input feature set with uniform dimensionality and higher information density is generated and output to the subsequent prediction model.

[0030] Therefore, by classifying and defining multi-source risk data, the coverage dimensions of risk data are clarified, laying a systematic data foundation for comprehensively depicting policy risks. By extracting risk features and time features, the original heterogeneous business data is transformed into quantifiable risk representation dimensions, while taking into account both static risk levels and dynamic time fluctuation patterns. Through cross-processing of industry and occupational features, the joint risk interaction effects that cannot be reflected by single features are explicitly mined, strengthening the expressive power of high-dimensional risk patterns. Finally, by dimensionality reduction processing to compress the redundancy of high-dimensional features, core risk information is preserved while reducing the computational complexity and overfitting risk of subsequent models. This improves the comprehensiveness and accuracy of risk characterization from the feature processing level, providing high-quality input assurance for the accuracy of subsequent loss prediction.

[0031] In some embodiments, such as Figure 4 As shown, multi-source risk data includes at least policy core data, personnel exposure data, and external supplementary data. Step S11, which involves extracting risk characteristics and time characteristics from the policy core data, personnel exposure data, and external supplementary data, specifically includes the following steps S111 to S116: S111: Obtain the occupational baseline score for each target occupation from the personnel exposure data and the occupational accident rate coefficient for each target occupation from the external supplementary data.

[0032] Among them, the occupational benchmark score refers to the benchmark quantitative score that reflects the inherent danger level of different occupational positions, which is preset based on a unified occupational risk classification standard. The higher the occupational danger level, the higher the corresponding benchmark score. The occupational accident rate coefficient refers to the ratio of the accident rate of the target occupation's industry to the average accident rate of the entire industry. It is used to characterize the weighted impact coefficient of the overall risk level of the industry on the risk of a single occupation.

[0033] Specifically, the risk classification corresponding to each target occupation is matched from the personnel exposure data, and the preset occupational benchmark score is retrieved; at the same time, the accident rate data of the industry to which each target occupation belongs is extracted from the external supplementary data, the ratio of the industry accident rate to the average accident rate of the entire industry is calculated, and the occupational accident rate coefficient corresponding to each target occupation is obtained, providing basic parameters for subsequent risk scoring calculation.

[0034] S112: The occupational risk score is obtained by weighting and integrating the baseline scores of each target occupation according to the occupational accident rate coefficient.

[0035] Among them, the occupational risk score refers to the comprehensive occupational risk quantification value obtained after weighting by industry accident rates, which is used to characterize the overall risk level of the positions of employees in insured companies.

[0036] Specifically, the risk score for each target occupation is calculated by weighting the occupational accident rate coefficient corresponding to each target occupation and performing an occupational benchmark score. If the insured company includes multiple target occupations, the risk scores for each single occupation are further weighted and summed according to the proportion of employees in each occupation to the total number of employees, and finally the overall occupational risk score of the target company is obtained.

[0037] S113: Calculate the ratio of the deductible amount to the insured amount in the core policy data to obtain the deductible coverage ratio.

[0038] Among them, the deductible refers to the threshold amount that the insurance company is exempt from paying out as stipulated in the policy terms, i.e., the deductible. Losses below this amount are borne by the insured company. The insurance limit refers to the maximum payout limit for a single accident or single person as stipulated in the policy terms, which is the upper limit of the risk borne by the insurance company. The deductible coverage ratio refers to the ratio of the deductible to the corresponding insurance limit, which is used to quantify the degree to which the deductible clause offsets the overall payout risk. The higher the ratio, the lower the payout risk for the insurance company.

[0039] Specifically, the deductible amount stipulated in the policy and the corresponding insurance amount are extracted from the core data of the policy. The ratio of the deductible amount to the insurance amount is calculated to generate the deductible coverage ratio indicator, which transforms the risk adjustment effect of the deductible clause into a quantifiable numerical feature.

[0040] S114: Calculate the ratio of the historical claims amount to the historical premium of the target company corresponding to the policy to be predicted in the external supplementary data over a historical period to obtain the historical loss ratio.

[0041] Among them, the target enterprise refers to the insured entity corresponding to the policy to be predicted, which is the insured object of risk assessment; the historical period refers to the preset historical data statistical time window, which is used to limit the statistical scope of historical claims and accident data; the historical claims amount refers to the total cumulative claims amount of all claims that have occurred within the preset historical period; the historical premium refers to the total cumulative premium actually received by the target enterprise for the corresponding policy within the preset historical period; and the historical loss ratio refers to the ratio of historical claims amount to historical premium, which is used to directly reflect the actual risk performance and claims level of the insured enterprise in the past.

[0042] Specifically, all claims records and corresponding premium records of the target company within a preset historical period are selected from external supplementary data. The cumulative historical claims amount and cumulative historical premium are calculated separately, and the ratio between the two is calculated to obtain the historical loss ratio reflecting the company's historical risk performance. If the target company has no historical data, the average loss ratio of companies of the same size in the same industry is used as the filler value.

[0043] S115: The occupational risk score, exemption coverage rate, and historical loss ratio are identified as risk characteristics.

[0044] Among them, the three quantitative indicators calculated above—occupational risk score, exemption coverage rate, and historical loss ratio—are jointly determined as risk characteristics, serving as core numerical features characterizing the static risk level of the policy, and are included in the feature set for subsequent model input.

[0045] S116: Determine the time characteristics based on the target company's historical accident count and the time period of each historical accident in the external supplementary data.

[0046] Among them, the number of historical accidents refers to the total number of accidents that meet the compensation conditions that occurred in the target company within a preset historical period; the occurrence period refers to the time cycle attribute of the historical accidents, including time dimensions such as months and quarters, which are used to reflect the time fluctuation pattern of risk.

[0047] Specifically, the historical number of accidents of the target enterprise and the corresponding time period of each accident are extracted from external supplementary data. The periodic attributes of the time period are extracted and numerically encoded to generate time features that can reflect the time pattern of accident occurrence, which are used to capture the seasonal and periodic fluctuations of risk.

[0048] In this way, by weighting occupational risk scores by overlaying industry accident rate coefficients onto occupational benchmark scores, a quantitative integration of inherent occupational risks and overall industry risks is achieved, which is more in line with the pattern of industry and job risk linkage in real business. By waiving coverage limits, policy deductible clauses are transformed into quantitative risk characteristics, which intuitively reflects the regulatory role of deductible settings on payout risk. Historical payout rates directly quantify the company's past actual risk performance, providing a highly correlated historical reference for loss prediction. At the same time, by extracting time features from the historical accident occurrence periods, the seasonal fluctuation patterns of risk are effectively captured. The entire feature extraction logic is deeply aligned with the business rules of liability insurance, and the generated features have clear business meanings and significant risk differentiation, improving the accuracy and business rationality of loss prediction results from the feature construction level.

[0049] S20: Input the input feature set into the trained payout frequency prediction model and the trained payout intensity prediction model respectively to predict payout frequency and payout intensity, and output the expected payout frequency and expected payout intensity respectively.

[0050] Among them, the claim frequency prediction model refers to a pre-trained deep learning model specifically designed to predict the frequency of claim incidents per unit of risk exposure, outputting a quantitative indicator representing the probability of incident occurrence; the claim intensity prediction model refers to a pre-trained deep learning model specifically designed to predict the level of loss in a single claim incident, outputting a quantitative indicator representing the degree of loss in a single incident; the expected claim frequency refers to the frequency of expected claim incidents within the policy coverage period per unit of risk exposure, and is a core indicator for measuring the probability of risk occurrence; the expected claim intensity refers to the level of expected claim amount when a single claim incident occurs, and is a core indicator for measuring the severity of loss in a single incident.

[0051] Specifically, a standardized set of input features is simultaneously input into two pre-trained deep learning prediction models. The payout frequency prediction model infers and calculates based on the counting characteristics of accidents, and outputs the corresponding expected payout frequency value. The payout intensity prediction model infers and calculates based on the monetary loss characteristics of a single accident, and outputs the corresponding expected payout intensity value. The two models operate independently and are executed in parallel, respectively outputting two core indicators representing the probability of risk occurrence and the degree of risk loss.

[0052] In some embodiments, such as Figure 5 As shown, in step S20, the input feature set is input into the trained payout frequency prediction model and the trained payout intensity prediction model respectively to predict the payout frequency and payout intensity, and the expected payout frequency and expected payout intensity are output respectively. Specifically, this includes the following steps S21 to S24: S21: Input the input feature set into the trained payout frequency prediction model, perform multi-layer fully connected forward computation, and output zero-inflated Poisson distribution parameters.

[0053] Among them, multi-layer fully connected forward computation refers to the computational process in a deep neural network in which the input feature vector passes through multiple layers of fully connected neurons in sequence, performing linear transformations and nonlinear activation operations layer by layer, and passing the feature information from the input layer to the output layer. It is the core operation method for deep learning to mine nonlinear correlations of features. The zero-inflation Poisson distribution parameter refers to the core parameter used to fully characterize the zero-inflation Poisson distribution. It includes at least the zero-inflation probability parameter and the Poisson strength parameter. The former represents the probability of zero payout during the policy coverage period, and the latter represents the average frequency of accidents occurring under non-zero payout conditions.

[0054] Specifically, the standardized input feature set is input into the input layer of the payout frequency prediction model, and then passed through a multi-layer fully connected network for progressive calculation. After each layer performs a linear weighted transformation on the input features, a nonlinear mapping is introduced through an activation function to refine higher-order risk features layer by layer. Finally, the mapping calculation is completed in the output layer, and the two sets of core parameters corresponding to the zero-inflated Poisson distribution are output, thus completing the transformation from risk features to frequency distribution patterns.

[0055] S22: Calculate the expected payout frequency based on the zero-inflated Poisson distribution parameters.

[0056] The mathematical expectation calculation formula of the zero-inflated Poisson distribution is called, and the zero-inflated probability parameter and Poisson strength parameter output by the model are substituted to calculate the mathematical expectation value of the frequency of compensation events under a unit risk exposure scale, that is, the final expected compensation frequency.

[0057] S23: Input the input feature set into the trained payout intensity prediction model, and sequentially perform local feature interaction extraction, temporal evolution feature extraction, and risk feature weighting processing to output log-normal distribution parameters.

[0058] Among them, local feature interaction extraction refers to the feature processing process of mining the combination rules of strongly correlated risk factors from input features and extracting high-order risk information generated by the local linkage of different risk factors, which is used to capture the joint risk effect that cannot be reflected by a single feature; temporal evolution feature extraction refers to the feature processing process of mining the long-term trend, short-term fluctuation and periodic pattern of risk level changes over time from feature sequences with time attributes, which is used to reflect the dynamic change characteristics of risk; risk feature weighting processing refers to the process of automatically assigning differentiated weights to different features according to the contribution of each risk feature to the payout intensity, amplifying the influence of core features with high contribution and weakening the interference of low-correlation features; log-normal distribution parameters refer to the core parameters used to completely characterize the shape of the log-normal distribution, including at least the logarithmic mean parameter and the logarithmic variance parameter; the payout amount follows this normal distribution after taking the natural logarithm, which can accurately characterize the right-skewed heavy-tailed data characteristics of the payout amount.

[0059] Specifically, after inputting the feature set into the payout intensity prediction model, the model first performs local feature interaction extraction to uncover the combined risk patterns of strongly correlated risk factors such as industry, region, and occupation, generating higher-order combined risk features. Next, it performs time-series evolution feature extraction to learn the long-term upward trend and seasonal fluctuations of payout intensity from historical time-series data. Then, it performs risk feature weighting processing to automatically calculate the contribution of each feature to payout intensity and assign corresponding weights, strengthening the influence weight of core risk factors. Finally, the model obtains the core parameters corresponding to the log-normal distribution through output layer mapping calculation.

[0060] S24: Calculate the expected payout intensity based on the log-normal distribution parameters.

[0061] The method involves calling the mathematical expectation calculation formula of the log-normal distribution, substituting the logarithmic mean parameter and logarithmic variance parameter output by the model, and calculating the mathematical expectation value of the loss amount corresponding to a single compensation accident, which is the final expected compensation intensity.

[0062] Therefore, a customized inference architecture is adopted, which outputs distribution parameters from a deep learning network and then calculates the expected value based on the statistical distribution, to replace the traditional prediction method that directly outputs a single value. Among them, the payout frequency prediction uses a multi-layer fully connected network combined with zero-inflated Poisson distribution parameters to accurately adapt to the characteristics of count data with a high proportion of zero payout frequencies. The payout intensity prediction uses a progressive processing of local interactive extraction, temporal evolution extraction, and feature weighting, combined with log-normal distribution parameters to effectively adapt to the right-skewed heavy-tailed data characteristics of payout amounts. The two branches are customized to model loss indicators with two different distribution characteristics, which not only give full play to the nonlinear fitting advantages of deep learning, but also ensure that the output results conform to the statistical laws of insurance business through statistical distribution constraints. This significantly improves the prediction accuracy of the two core indicators of payout frequency and payout intensity, and provides a reliable intermediate calculation basis for subsequent overall loss calculation.

[0063] In some embodiments, the trained payout intensity prediction model includes a convolutional neural network layer, a long short-term memory network layer, and an attention layer. The input feature set is input into the trained payout intensity prediction model, and local feature interaction extraction, temporal evolution feature extraction, and risk feature weighting are performed sequentially to output log-normal distribution parameters. Specifically, the steps include: The input feature set is input into the trained payout intensity prediction model. Local feature interaction is extracted through the convolutional neural network layer, temporal evolution feature is extracted through the long short-term memory network layer, and risk feature weighting is performed through the attention layer to output normal distribution parameters.

[0064] The convolutional neural network layer in the payout intensity prediction model is used to perform local feature interaction extraction. It uses one-dimensional convolution operations to slide and scan the ordered input feature sequence, automatically mining the combination and linkage patterns of strongly correlated risk factors, and generating feature data representing high-order combination risks. The long short-term memory network layer in the payout intensity prediction model is used to perform temporal evolution feature extraction. It has a gated structure with forget gate, input gate, and output gate, which can retain long-term trend information of risk changes, filter outdated and invalid historical information, and accurately capture the dynamic pattern of payout intensity evolution over time. The attention layer in the payout intensity prediction model is used to perform risk feature weighting. It generates corresponding weights by automatically calculating the contribution of each feature dimension to the payout intensity prediction result, amplifying the influence weight of core risk features, weakening the interference of low-correlation features, and achieving focused calculation of key risk factors.

[0065] Specifically, the input feature set is arranged into a feature sequence according to feature attributes and then input into the trained payout intensity prediction model. First, a sliding convolution operation is performed on the feature sequence through a convolutional neural network layer to mine the combined correlation of adjacent risk factors segment by segment, extracting local joint risk features of dimensions such as industry and region, occupation and insured amount, and completing the local feature interaction extraction. Then, the output feature sequence is fed into a long short-term memory network layer, and the time-series feature information is filtered and updated through a gating structure to retain the long-term evolution trend of payout intensity and capture short-term seasonal fluctuation patterns, thus completing the time-series evolution feature extraction. Subsequently, the feature vector fused with time-series information is input into the attention layer, which automatically calculates the contribution weight of each feature dimension and performs a weighted operation to strengthen the influence of high-risk features on the output results. Finally, the output layer maps and calculates to output the log-normal distribution parameters.

[0066] Therefore, through a hierarchical combination architecture of convolutional neural network layers, long short-term memory network layers, and attention layers, progressive processing is achieved for local combined risk mining, temporal dynamic pattern capture, and core risk factor focusing, respectively. Compared with single-structure prediction models, it can extract multi-dimensional risk features related to payout intensity more comprehensively and efficiently. It not only fully explores the joint risk effects of static attributes such as industry, region, and occupation, but also accurately captures the dynamic trend of payout level changes over time. At the same time, the attention mechanism automatically anchors high-impact risk factors, reduces the interference of redundant features, and effectively improves the model's fitting accuracy to the heavy-tailed distribution characteristics of payout intensity, providing a more reliable intermediate calculation basis for subsequent accurate risk loss measurement.

[0067] S30: Perform distribution verification on the expected payout frequency and expected payout intensity to obtain the verification results.

[0068] Distribution verification refers to the process of verifying the statistical rationality of the output of a deep learning model based on the classical statistical distribution of insurance losses. It is used to determine whether the prediction results conform to the inherent statistical characteristics of insurance business. The verification result refers to the verification conclusion output by the distribution verification process. It is used to quantitatively characterize the degree of fit and deviation level between the prediction results and the corresponding statistical distribution, and serves as the basis for subsequent deviation calibration.

[0069] Specifically, the expected payout frequency output by the model is substituted into the corresponding payout frequency statistical distribution model to verify whether it conforms to the inherent statistical law of insurance payout frequency; at the same time, the expected payout intensity is substituted into the corresponding payout intensity statistical distribution model to verify whether it conforms to the distribution characteristics of insurance payout amount; the degree of deviation between the predicted value and the theoretical value of the statistical distribution is calculated respectively, and the corresponding frequency verification results and intensity verification results are generated as the basis for subsequent calibration.

[0070] In some embodiments, such as Figure 6 As shown, step S30, which involves performing a distribution verification on the expected payout frequency and expected payout intensity to obtain the verification result, specifically includes the following steps S31 to S32: S31: Substitute the expected payout frequency into the zero-inflated Poisson distribution, verify the statistical rationality of the frequency prediction through the information criterion, and obtain the frequency verification result.

[0071] Among them, the zero-inflated Poisson distribution is a mixed probability distribution suitable for count data with a significantly high proportion of zero values. It is composed of a zero-state determination part and a standard Poisson distribution part, which can accurately match the distribution characteristics of the frequency of liability insurance claims. The information criterion is a general evaluation standard used to quantify the goodness of fit of the statistical model to the sample data. It judges the rationality of the distribution model by comprehensively balancing the fitting accuracy and model simplicity, and is a quantitative basis for verifying the statistical compliance of frequency prediction. The frequency verification result is the verification conclusion data output after performing distribution verification on the expected payment frequency. It includes quantitative information such as the degree of fit between the predicted distribution and the theoretical distribution and the level of deviation, and is used to judge the statistical rationality of the frequency prediction results.

[0072] Specifically, the distribution parameters corresponding to the expected payout frequency output by the payout frequency prediction model are substituted into the zero-inflated Poisson distribution to construct the corresponding theoretical frequency distribution model; the information criterion is called to calculate the quantified value of the fit of the theoretical distribution to the payout frequency sample data; based on the quantified value, it is determined whether the frequency prediction result conforms to the inherent statistical law of insurance payout frequency; finally, a frequency verification result containing fitting deviation and reasonableness conclusions is generated.

[0073] S32: Substitute the expected compensation intensity into the log-normal distribution, verify the rationality of the intensity distribution through the goodness-of-fit test, and obtain the intensity verification result.

[0074] Among them, the log-normal distribution is a continuous probability distribution. Its random variable follows a normal distribution after being transformed by the natural logarithm. It can effectively characterize the right-skewed and heavy-tailed data characteristics of the compensation amount and is suitable for the distribution pattern of extreme large losses. The goodness-of-fit test is a statistical test method used to test whether the actual distribution of the sample data is consistent with the specified theoretical probability distribution. It determines the fitting effect by quantifying the cumulative distribution deviation between the two. The strength verification result is the verification conclusion data output after performing distribution verification on the expected compensation strength. It includes quantitative information such as the overall deviation between the strength prediction distribution and the theoretical distribution and the degree of tail fit, and is used to judge the reasonableness of the distribution of the strength prediction result.

[0075] Specifically, the distribution parameters corresponding to the expected payout intensity output by the payout intensity prediction model are substituted into the log-normal distribution to construct the corresponding theoretical intensity distribution model. The cumulative probability deviation between the actual distribution of the predicted payout intensity and the theoretical log-normal distribution is calculated through a goodness-of-fit test. The tail fitting effect of high loss quantiles is checked in detail to determine the rationality of the distribution of the intensity prediction results. Finally, an intensity verification result containing the overall deviation and tail fitting conclusions is generated.

[0076] Therefore, for two different data characteristics of indicators, namely payout frequency and payout intensity, appropriate theoretical statistical distributions and verification methods are used to perform rationality verification. The output range of the deep learning model is constrained by the mature classical statistical laws in the insurance field. This can effectively identify abnormal prediction results that deviate from business laws due to overfitting of pure data-driven models, and provide clear quantitative judgment basis for subsequent deviation calibration. From the perspective of verification mechanism, it ensures that the prediction results are in line with the statistical logic of insurance actuarial science, significantly improves the reliability and business adaptability of loss prediction results, and avoids the prediction distortion problem that is prone to occur in unconstrained deep learning models.

[0077] S40: Based on the verification results, perform deviation calibration on the expected payout frequency and expected payout intensity respectively to obtain the calibrated frequency distribution parameters and intensity distribution parameters.

[0078] Among them, deviation calibration refers to the process of correcting and adjusting the prediction results when the distribution verification shows that the prediction results deviate from the theoretical values ​​of the statistical distribution beyond the preset range. The purpose is to ensure that the prediction results conform to the statistical logic of insurance actuarial science. Frequency distribution parameters refer to the core parameters used to describe the statistical distribution corresponding to the claim frequency, which determine the distribution pattern and statistical characteristics of the claim frequency. Intensity distribution parameters refer to the core parameters used to describe the statistical distribution corresponding to the claim intensity, which determine the distribution pattern and statistical characteristics of the claim intensity.

[0079] Specifically, based on the deviation results obtained from the distribution verification, it is determined whether the deviation between the predicted value and the expected value of the statistical distribution exceeds a preset threshold. If the deviation is within the threshold range, the corresponding frequency distribution parameters and intensity distribution parameters are directly generated based on the predicted value. If the deviation exceeds the preset threshold, the predicted value is weighted and corrected according to the preset calibration rules. After correction, calibrated frequency distribution parameters and intensity distribution parameters are generated to ensure that the distribution parameters meet the statistical rationality of insurance actuarial calculation.

[0080] S50: Construct a composite loss distribution based on frequency distribution parameters and intensity distribution parameters, and calculate the expected value of the loss distribution.

[0081] Among them, the composite loss distribution refers to the composite probability distribution constructed by combining the payout frequency distribution and the payout intensity distribution. It can fully describe the probability distribution law of the total policy loss and cover all scenarios of loss from low frequency and small amount to high frequency and large amount. The expected value of the loss distribution refers to the mathematical expectation corresponding to the composite loss distribution, that is, the expected loss amount under a unit of risk exposure, which is the core benchmark value characterizing the average risk loss level.

[0082] Specifically, the two independent distributions are combined using the calibrated frequency distribution parameter as the distribution parameter for the frequency dimension and the calibrated intensity distribution parameter as the distribution parameter for the loss degree dimension to construct a composite loss distribution corresponding to the total policy loss. Based on the probability density function of this composite loss distribution, the expected value of the distribution, i.e., the expected loss amount per unit of risk exposure, is calculated.

[0083] S60: Determine the risk loss based on the expected value of the loss distribution and the risk exposure corresponding to the policy to be predicted.

[0084] Among them, risk exposure refers to a quantitative indicator that characterizes the scale of risk underwritten by the policy. It is used to measure the size of the risk exposure corresponding to the policy and is the calculation base connecting the unit risk indicator and the total loss of the policy. Risk loss refers to the final calculation result output by this method, that is, the expected total loss amount corresponding to the policy during the protection period.

[0085] Specifically, relevant indicators for calculating the risk scale are extracted from the basic data of the policy to be predicted, and the total risk exposure corresponding to the policy is calculated. The expected value of the loss distribution corresponding to the unit risk is multiplied by the total risk exposure to obtain the basic risk loss scale corresponding to the policy. Based on this, the tail risk characteristics of the distribution are combined to complete the final calculation and obtain the final risk loss result corresponding to the policy to be predicted.

[0086] In some embodiments, such as Figure 7 As shown, step S60, which involves determining the risk loss based on the expected value of the loss distribution and the risk exposure corresponding to the policy to be predicted, specifically includes the following steps S61 to S63: S61: Calculate the risk exposure corresponding to the policy to be predicted based on the total income of the personnel in the policy to be predicted and the coverage period.

[0087] Among them, total personnel income refers to the total salary income of all insured persons within the coverage of the policy to be predicted. It is the core base indicator for calculating the scale of policy risk and corresponds to the personnel size dimension of risk exposure. Coverage period refers to the effective duration of the insurance liability stipulated in the policy. It is the time dimension indicator for calculating the scale of policy risk. The longer the coverage period, the greater the time exposure of risk exposure. Risk exposure refers to the quantitative value of the overall risk exposure of the policy obtained by combining personnel income scale and coverage period. It is used to amplify the calculation result of unit risk to the actual risk scale of the entire policy and serves as a conversion bridge connecting unit risk indicators and total policy loss.

[0088] Specifically, the total income of the insured persons and the coverage period stipulated in the policy are extracted from the core data of the policy to be predicted; the coverage period is converted into a standardized period coefficient according to a unified time standard; and the total income of the insured persons is multiplied by the coverage period coefficient to calculate the total risk exposure of the policy, thus completing the quantitative accounting of the overall risk scale of the single policy.

[0089] S62: Multiply the expected value of the loss distribution by the amount of risk exposure to obtain the basic risk loss value.

[0090] Among them, the basic risk loss value refers to the expected loss amount of the policy calculated based on the average risk level. It only covers the expected value of the loss distribution and is the benchmark value for risk loss calculation, corresponding to the regular expected loss scale of the policy.

[0091] Specifically, the expected value of unit risk loss output by the composite loss distribution is used as a standardized benchmark value. This value is then multiplied with the total risk exposure calculated above. The average loss index at the unit scale is converted into the average expected total loss for the entire policy, resulting in a basic risk loss value that only covers the average risk level. This serves as the core benchmark for loss calculation.

[0092] S63: Calculate the risk surcharge based on the preset loss quantile of the composite loss distribution, and add it to the basic risk loss value to obtain the risk loss.

[0093] Among them, the preset loss quantile refers to the loss quantile value corresponding to the high confidence level in the composite loss distribution. It is used to characterize the risk level of extreme large losses at the tail and is the core reference indicator for measuring extreme risks and setting aside risk redundancy. The risk additional limit refers to the additional loss limit calculated based on the preset loss quantile to cover extreme risks at the tail. It is used to make up for the problem that the basic risk loss value is insufficient to cover extreme large compensation and improve the risk coverage of the calculation results.

[0094] Specifically, the extreme loss quantiles of the preset levels are extracted from the constructed composite loss distribution. The difference between the extreme loss quantiles and the expected value of the loss distribution is calculated, and the corresponding risk surcharge is calculated in combination with the risk surcharge rules. The risk surcharge is then added to the basic risk loss value so that the final calculation result covers both the conventional average expected loss and reserves risk redundancy for extreme large payouts. Finally, the complete risk loss calculation result is output.

[0095] In this way, by calculating the total risk exposure through the total income of personnel and the coverage period, the accurate conversion from standardized unit risk indicators to the actual risk scale of a single policy is achieved, ensuring the matching of the calculation results with the actual underwriting scale of the policy. The basic risk loss value is obtained by multiplying the expected value of loss distribution by the risk exposure, which anchors the average expected loss level of the policy. On this basis, a risk additional amount calculated based on the extreme loss quantile is superimposed, which effectively makes up for the deficiency of insufficient coverage of extreme risks at the tail end by the single expected value calculation. This makes the final loss calculation result take into account both the accuracy of average loss and the ability to cover extreme risks, deeply aligning with the business logic of insurance risk measurement, and significantly improving the completeness and risk reliability of the loss calculation result.

[0096] In this embodiment of the invention, multi-source risk data corresponding to the policy to be predicted is acquired, and feature engineering is performed on the multi-source risk data to generate an input feature set. The input feature set is then input into a trained claim frequency prediction model and a trained claim intensity prediction model to predict claim frequency and claim intensity, respectively, outputting the expected claim frequency and expected claim intensity. Distribution verification is performed on the expected claim frequency and expected claim intensity to obtain verification results. Based on the verification results, deviation calibration is performed on the expected claim frequency and expected claim intensity to obtain calibrated frequency distribution parameters and intensity distribution parameters. A composite loss distribution is constructed based on the frequency distribution parameters and intensity distribution parameters, and the expected value of the loss distribution is calculated. Based on the expected value of the loss distribution and the risk exposure corresponding to the policy to be predicted, the risk loss is determined. This invention, by acquiring multi-source risk data of the policy to be predicted and performing feature engineering, breaks through the limitations of traditional solutions. By relying solely on a few basic risk indicators, this approach integrates and standardizes emerging multi-dimensional risk factors, addressing the shortcomings of incomplete risk characterization at the data input level. Secondly, it employs a parallel prediction architecture using dual-branch deep learning models for payout frequency and payout intensity. Leveraging the non-linear feature mining capabilities of deep learning, it learns the implicit correlation between risk factors and losses. Simultaneously, it models data with different distribution characteristics for accident frequency and payout amount, enhancing the model's adaptability to volatile data and small-sample scenarios, improving generalization performance in new scenarios. Finally, through a closed-loop constraint mechanism constructed using distribution verification, bias calibration, and composite loss distribution, it verifies and corrects the deep learning output results based on the inherent statistical laws of insurance losses, suppressing the overfitting risk of purely data-driven models and ensuring that the prediction results conform to the statistical rationality of actuarial science. Thus, it addresses the shortcomings of existing technologies layer by layer from three levels: data dimension, modeling capability, and result constraints, achieving an overall improvement in the accuracy and reliability of loss prediction.

[0097] In one embodiment, a loss prediction device is provided, which corresponds one-to-one with the loss prediction methods described in the above embodiments. For example... Figure 8As shown, the loss prediction device includes a feature processing unit 801, an input unit 802, a verification unit 803, a calibration unit 804, a calculation unit 805, and a determination unit 806. Detailed descriptions of each functional module are as follows: Feature processing unit 801 is used to acquire multi-source risk data corresponding to the policy to be predicted, perform feature engineering processing on the multi-source risk data, and generate an input feature set; The input unit 802 is used to input the input feature set into the trained payout frequency prediction model and the trained payout intensity prediction model respectively to predict the payout frequency and payout intensity, and output the expected payout frequency and expected payout intensity respectively. The verification unit 803 is used to perform distribution verification on the expected payout frequency and the expected payout intensity to obtain the verification result. The calibration unit 804 is used to perform deviation calibration on the expected payout frequency and the expected payout intensity according to the verification result, so as to obtain the calibrated frequency distribution parameters and intensity distribution parameters. The calculation unit 805 is used to construct a composite loss distribution based on the frequency distribution parameters and the intensity distribution parameters, and to calculate the expected value of the loss distribution. The determining unit 806 is used to determine the risk loss based on the expected value of the loss distribution and the risk exposure corresponding to the policy to be predicted.

[0098] In one embodiment, multi-source risk data includes at least core policy data, personnel exposure data, and external supplementary data; Feature processing unit 801 specifically includes: The extraction sub-unit is used to extract risk features and time features from the core policy data, personnel exposure data, and external supplementary data. The feature crossover subunit is used to perform feature crossover processing on the target industry of the policy core data and the target occupation of the personnel exposure data to obtain combined features; The splicing subunit is used to splice the risk feature, the time feature, and the combined feature to obtain a spliced ​​feature set; The dimensionality reduction subunit is used to perform dimensionality reduction processing on the spliced ​​feature set to obtain the input feature set.

[0099] In one embodiment, the extracted sub-unit is specifically used for: Obtain the occupational baseline score for each target occupation in the personnel exposure data and the occupational accident rate coefficient for each target occupation in the external supplementary data; The occupational risk score is obtained by weighting and integrating the baseline score of each target occupation according to the occupational accident rate coefficient of each occupation. Calculate the ratio of the deductible amount to the insured amount in the core policy data to obtain the deductible coverage rate; The historical loss ratio is obtained by calculating the ratio of the historical claims amount to the historical premium of the target enterprise corresponding to the policy to be predicted in the external supplementary data over a historical period. The occupational risk score, the exemption coverage rate, and the historical payout rate are identified as risk characteristics. Based on the target company's historical accident count and the time period of each historical accident in the external supplementary data, time characteristics are determined.

[0100] In one embodiment, the input unit 802 specifically includes: The first input subunit is used to input the input feature set into the trained payout frequency prediction model, perform multi-layer fully connected forward computation, and output zero-inflated Poisson distribution parameters. The first calculation subunit is used to calculate the expected payout frequency based on the zero-inflated Poisson distribution parameters; The second input subunit is used to input the input feature set into the trained payout intensity prediction model, and sequentially perform local feature interaction extraction, temporal evolution feature extraction, and risk feature weighting processing to output log-normal distribution parameters. The second calculation subunit is used to calculate the expected compensation intensity based on the log-normal distribution parameters.

[0101] In one embodiment, the trained payout intensity prediction model includes a convolutional neural network layer, a long short-term memory network layer, and an attention layer; The second input subunit is specifically used for: The input feature set is input into the trained payout intensity prediction model. Local feature interaction is extracted through the convolutional neural network layer, temporal evolution feature is extracted through the long short-term memory network layer, and risk feature weighting is performed through the attention layer to output normal distribution parameters.

[0102] In one embodiment, the verification unit 803 specifically includes: The first verification subunit is used to substitute the expected payout frequency into a zero-inflated Poisson distribution, verify the statistical rationality of the frequency prediction through information criteria, and obtain the frequency verification result. The second verification subunit is used to substitute the expected compensation intensity into a log-normal distribution, verify the rationality of the intensity distribution through a goodness-of-fit test, and obtain the intensity verification result.

[0103] In one embodiment, the determining unit 806 specifically includes: The third calculation subunit is used to calculate the risk exposure corresponding to the policy to be predicted based on the total income of the personnel in the policy to be predicted and the coverage period. The fourth calculation subunit is used to multiply the expected value of the loss distribution by the risk exposure to obtain the basic risk loss value; The fifth calculation subunit is used to calculate the risk surcharge based on the preset loss quantile of the composite loss distribution, and then add it to the basic risk loss value to obtain the risk loss.

[0104] This invention provides a loss prediction device. A feature processing unit 801 acquires multi-source risk data corresponding to the policy to be predicted, performs feature engineering on the multi-source risk data, and generates an input feature set. An input unit 802 inputs the input feature set into a trained claim frequency prediction model and a trained claim intensity prediction model to predict the claim frequency and claim intensity, respectively, and outputs the expected claim frequency and expected claim intensity. A verification unit 803 performs distribution verification on the expected claim frequency and expected claim intensity to obtain verification results. A calibration unit 804 performs deviation calibration on the expected claim frequency and expected claim intensity based on the verification results to obtain calibrated frequency distribution parameters and intensity distribution parameters. A calculation unit 805 constructs a composite loss distribution based on the frequency distribution parameters and intensity distribution parameters and calculates the expected value of the loss distribution. A determination unit 806 determines the risk loss based on the expected value of the loss distribution and the risk exposure corresponding to the policy to be predicted.

[0105] By acquiring multi-source risk data from the policies to be predicted and performing feature engineering, this approach overcomes the dimensional limitations of traditional solutions that rely solely on a few basic risk indicators. It integrates and standardizes multi-dimensional emerging risk factors, addressing the shortcomings of incomplete risk characterization at the data input level. Secondly, it employs a parallel prediction architecture using dual-branch deep learning models for payout frequency and payout intensity. Leveraging the non-linear feature mining capabilities of deep learning, it learns the implicit correlation between risk factors and losses. Simultaneously, it models data with different distribution characteristics for accident frequency and payout amount, enhancing the model's adaptability to volatile data and small-sample scenarios, and improving generalization performance in new scenarios. Finally, through a closed-loop constraint mechanism constructed using distribution verification, bias calibration, and composite loss distribution, it verifies and corrects the deep learning output results based on the inherent statistical laws of insurance losses, suppressing the overfitting risk of purely data-driven models and ensuring that the prediction results conform to the statistical rationality of actuarial science. Thus, it addresses the shortcomings of existing technologies layer by layer from three levels: data dimension, modeling capability, and result constraints, achieving an overall improvement in the accuracy and reliability of loss prediction.

[0106] Specific limitations regarding the loss prediction device can be found in the limitations of the loss prediction method described above, and will not be repeated here. Each module in the aforementioned loss prediction device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0107] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 9 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external clients via a network connection. When the computer program is executed by the processor, it implements the functions or steps of a loss prediction method on the server side.

[0108] In one embodiment, a computer device is provided, which may be a client, and its internal structure diagram may be as follows: Figure 10 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with an external server via a network connection. When the computer program is executed by the processor, it implements the functions or steps of a loss prediction method on the client side.

[0109] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps: Obtain multi-source risk data corresponding to the policy to be predicted, perform feature engineering processing on the multi-source risk data, and generate an input feature set; The input feature set is fed into the trained payout frequency prediction model and the trained payout intensity prediction model respectively to predict payout frequency and payout intensity, and the expected payout frequency and expected payout intensity are output respectively. The expected payout frequency and the expected payout intensity are subjected to distribution verification to obtain the verification results; Based on the verification results, deviation calibration is performed on the expected payout frequency and the expected payout intensity to obtain the calibrated frequency distribution parameters and intensity distribution parameters; A composite loss distribution is constructed based on the frequency distribution parameters and the intensity distribution parameters, and the expected value of the loss distribution is calculated. The risk loss is determined based on the expected value of the loss distribution and the risk exposure corresponding to the policy to be predicted.

[0110] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor: Obtain multi-source risk data corresponding to the policy to be predicted, perform feature engineering processing on the multi-source risk data, and generate an input feature set; The input feature set is fed into the trained payout frequency prediction model and the trained payout intensity prediction model respectively to predict payout frequency and payout intensity, and the expected payout frequency and expected payout intensity are output respectively. The expected payout frequency and the expected payout intensity are subjected to distribution verification to obtain the verification results; Based on the verification results, deviation calibration is performed on the expected payout frequency and the expected payout intensity to obtain the calibrated frequency distribution parameters and intensity distribution parameters; A composite loss distribution is constructed based on the frequency distribution parameters and the intensity distribution parameters, and the expected value of the loss distribution is calculated. The risk loss is determined based on the expected value of the loss distribution and the risk exposure corresponding to the policy to be predicted.

[0111] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0112] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0113] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0114] It should be noted that if any AI models, software tools, or components not belonging to this company appear in the embodiments of this invention, they are merely illustrative examples and do not represent actual use. All user personal information involved in the embodiments of this invention is authorized (with the knowledge and consent) by the relevant parties or fully authorized by all parties, and the executing entity may obtain it through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with relevant laws and regulations and do not violate public order and good morals.

[0115] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A loss prediction method, characterized in that, include: Obtain multi-source risk data corresponding to the policy to be predicted, perform feature engineering processing on the multi-source risk data, and generate an input feature set; The input feature set is input into the trained payout frequency prediction model and the trained payout intensity prediction model respectively to predict payout frequency and payout intensity, and output the expected payout frequency and expected payout intensity respectively. The expected payout frequency and the expected payout intensity are subjected to distribution verification to obtain the verification results; Based on the verification results, deviation calibration is performed on the expected payout frequency and the expected payout intensity to obtain the calibrated frequency distribution parameters and intensity distribution parameters; A composite loss distribution is constructed based on the frequency distribution parameters and the intensity distribution parameters, and the expected value of the loss distribution is calculated. The risk loss is determined based on the expected value of the loss distribution and the risk exposure corresponding to the policy to be predicted.

2. The loss prediction method according to claim 1, characterized in that, The multi-source risk data includes at least core policy data, personnel exposure data, and external supplementary data. The step of performing feature engineering on the multi-source risk data to generate an input feature set includes: Risk characteristics and time characteristics are extracted from the core policy data, personnel exposure data, and external supplementary data. The target industry of the core policy data and the target occupation of the personnel exposure data are subjected to feature cross-processing to obtain combined features; By concatenating the risk features, the time features, and the combined features, a concatenated feature set is obtained. The concatenated feature set is then subjected to dimensionality reduction processing to obtain the input feature set.

3. The loss prediction method according to claim 2, characterized in that, The extraction of risk characteristics and time characteristics from the core policy data, personnel exposure data, and external supplementary data includes: Obtain the occupational baseline score for each target occupation in the personnel exposure data and the occupational accident rate coefficient for each target occupation in the external supplementary data; The occupational risk score is obtained by weighting and integrating the baseline score of each target occupation according to the occupational accident rate coefficient of each occupation. Calculate the ratio of the deductible amount to the insured amount in the core policy data to obtain the deductible coverage rate; The historical loss ratio is obtained by calculating the ratio of the historical claims amount to the historical premium of the target enterprise corresponding to the policy to be predicted in the external supplementary data over a historical period. The occupational risk score, the exemption coverage rate, and the historical payout rate are identified as risk characteristics. Based on the target company's historical accident count and the time period of each historical accident in the external supplementary data, time characteristics are determined.

4. The loss prediction method according to claim 1, characterized in that, The step of inputting the input feature set into the trained payout frequency prediction model and the trained payout intensity prediction model respectively to predict payout frequency and payout intensity, and outputting the expected payout frequency and expected payout intensity respectively, includes: The input feature set is input into the trained payout frequency prediction model, and a multi-layer fully connected forward computation is performed to output zero-inflated Poisson distribution parameters. Calculate the expected payout frequency based on the zero-inflated Poisson distribution parameters; The input feature set is input into the trained payout intensity prediction model, and local feature interaction extraction, temporal evolution feature extraction, and risk feature weighting are performed sequentially to output log-normal distribution parameters. The expected payout intensity is calculated based on the log-normal distribution parameters.

5. The loss prediction method according to claim 4, characterized in that, The trained payout intensity prediction model includes a convolutional neural network layer, a long short-term memory network layer, and an attention layer. The input feature set is input into the trained payout intensity prediction model, which then sequentially performs local feature interaction extraction, temporal evolution feature extraction, and risk feature weighting processing to output log-normal distribution parameters, including: The input feature set is input into the trained payout intensity prediction model. Local feature interaction is extracted through the convolutional neural network layer, temporal evolution feature is extracted through the long short-term memory network layer, and risk feature weighting is performed through the attention layer to output normal distribution parameters.

6. The loss prediction method according to claim 1, characterized in that, The distribution verification of the expected payout frequency and the expected payout intensity to obtain the verification results includes: Substituting the expected payout frequency into a zero-inflated Poisson distribution, the statistical rationality of the frequency prediction is verified using the information criterion to obtain the frequency verification result; Substituting the expected compensation intensity into a log-normal distribution, the rationality of the intensity distribution is verified through a goodness-of-fit test, and the intensity verification result is obtained.

7. The loss prediction method according to claim 1, characterized in that, The step of determining the risk loss based on the expected value of the loss distribution and the risk exposure corresponding to the policy to be predicted includes: Based on the total income of the individuals in the policy to be predicted and the coverage period, calculate the risk exposure corresponding to the policy to be predicted; Multiply the expected value of the loss distribution by the risk exposure to obtain the basic risk loss value; The risk surcharge is calculated based on the preset loss quantile of the composite loss distribution and then added to the basic risk loss value to obtain the risk loss.

8. A loss prediction device, characterized in that, include: The feature processing unit is used to acquire multi-source risk data corresponding to the policy to be predicted, perform feature engineering processing on the multi-source risk data, and generate an input feature set. The input unit is used to input the input feature set into the trained payout frequency prediction model and the trained payout intensity prediction model respectively to predict the payout frequency and payout intensity, and output the expected payout frequency and expected payout intensity respectively. A verification unit is used to perform distribution verification on the expected payout frequency and the expected payout intensity to obtain verification results. The calibration unit is used to perform deviation calibration on the expected payout frequency and the expected payout intensity according to the verification result, so as to obtain the calibrated frequency distribution parameters and intensity distribution parameters. The calculation unit is used to construct a composite loss distribution based on the frequency distribution parameters and the intensity distribution parameters, and to calculate the expected value of the loss distribution; The determining unit is used to determine the risk loss based on the expected value of the loss distribution and the risk exposure corresponding to the policy to be predicted.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the loss prediction step as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the loss prediction steps as described in any one of claims 1 to 7.

Images