Short insurance dynamic pricing method based on AI large model

Abstract

The invention discloses a short insurance dynamic pricing method based on an AI large model, and the method comprises the steps: S1, constructing a multi-modal risk dynamic perception architecture, and outputting a unified dynamic risk vector with a fixed length; s2, establishing a real-time risk pricing adaptive mechanism based on the unified dynamic risk vector with the fixed length output in S1, and realizing a real-time pricing decision across insurance types; s3, designing an intelligent strategy generation and conflict resolution system, optimizing the real-time pricing decision across insurance types in the S2, and ensuring that the pricing decision achieves the optimal balance between market competitiveness and profit targets on the premise of meeting actuarial constraints and supervision requirements; and S4, constructing a complete decision tracing mechanism by adopting a multi-mode interpretable artificial intelligence technology, and generating an attribution report based on the data output from S1 to S3 so as to realize the conversion of the pricing decision from a black box to a white box. According to the patent technology, through innovative application of an AI large model, a new-generation intelligent pricing solution is provided for short-term insurance services.

Description

Technical Field

[0001] This application relates to the interdisciplinary field of artificial intelligence and insurance pricing, specifically to a dynamic pricing method for short-term insurance based on a large AI model. Background Technology

[0002] Short-term insurance dynamic pricing technology is undergoing a leapfrog development from experience-driven static pricing to AI-enabled real-time accurate pricing, with the core evolution path revolving around...

[0003] The research revolves around three main technical lines: risk quantification modeling, intelligent algorithm optimization, and multi-source data fusion. Its technological iteration logic and existing bottlenecks are as follows:

[0004] 1. Static Statistical Modeling Stage: Risk Quantification Based on Empirical Rules. Early short-term insurance pricing relied on actuarial rules and basic statistical models, with the core objective of achieving preliminary risk classification and pricing. In 2010, Smith J. proposed a multiple linear regression pricing model, using limited static factors such as age and occupation to determine the basic premium rate. However, this model could not capture the dynamic changes and differentiated characteristics of individual risk. In 2013, Zhang Wei used a generalized linear model (GLM) to optimize risk factor weights, improving pricing stability. However, the model was limited by pre-set statistical distribution assumptions and was difficult to adapt to unstructured data such as text and behavioral trajectories. In 2015, Li Ming used Bayesian networks to integrate historical claims data, strengthening the logic of risk quantification. However, this relied on manually-led feature engineering, and the pricing scheme update cycle was as long as six months to a year, making it unable to respond to market fluctuations and sudden risks. In 2017, Wang Hong proposed an interval grouping pricing model, which reduced computational costs by simplifying risk classification. However, it suffered from insufficient pricing fairness due to ignoring the heterogeneity of individual risk and could not meet personalized needs.

[0005] 2. Machine Learning Optimization Stage: Multi-Factor Accurate Risk Identification. With the expansion of data dimensions, machine learning algorithms have become a core tool for overcoming the limitations of traditional models, focusing on improving pricing accuracy and factor mining capabilities. In 2018, YuJiang used the Random Forest algorithm to integrate customer basic information and claims data, reducing pricing error by 8.3% compared to traditional models. However, in high-dimensional sparse data scenarios, feature redundancy easily occurs, affecting model performance. In 2020, Chen Yi enhanced the capture of feature interaction relationships through Gradient Boosting Tree (XGBoost), improving the granularity of risk identification. However, its performance is highly dependent on high-quality labeled data and sensitive to missing data and noise. In 2021, Robert's team combined Support Vector Machine (SVM) with a rule engine to achieve rapid dynamic adjustment of short-term insurance pricing. However, due to the "black box" nature of the model, interpretability was insufficient, facing regulatory compliance pressure. In 2022, Huang Bangju et al. proposed the LightGBM multi-factor pricing model, which significantly improved the pricing efficiency of small-amount short-term insurance. However, the data source was limited to the enterprise intranet and lacked the ability to adapt to sudden scenarios such as natural disasters and market changes in real time. In 2024, Wei Yongqiang and Xia Guangyu invented a method, device and storage medium for insurance product pricing and risk assessment based on lightGBM and neural networks for specific tasks. This patent mainly relies on structured data and preset rules for risk assessment. The data is single and cannot deeply understand unstructured data (such as medical text, vehicle damage images, voice recordings, etc.). The model update cycle is usually several weeks or even months, making it difficult to achieve true dynamic pricing.

[0006] 3. AI Large-Scale Model Integration Stage: Dynamic Risk Perception Based on Multi-Source Data. In recent years, AI large-scale models have become a core technology direction for dynamic pricing of short-term insurance due to their powerful heterogeneous data processing and feature learning capabilities. In 2023, ZhongAn Insurance team launched a health short-term insurance pricing scheme that deducts premiums based on steps, integrating wearable device health data with an open-source large-scale model. However, the model's adaptability is limited to a single scenario, and its generalization ability across insurance types (such as flight accident insurance and accident insurance) is insufficient. In 2024, Ping An Property & Casualty Insurance team used a large language model (LLM) to analyze driving behavior data collected by in-vehicle devices to achieve real-time dynamic adjustment of car insurance premiums, with a 25% increase in rates for high-risk users. However, the model has strict requirements for data anonymization and privacy protection, and large-scale parallel computing consumes a lot of hardware resources. In the same year, Li Xiao proposed a pricing framework that integrates STGNN and large-scale models, optimizing the timeliness of risk assessment through spatiotemporal sequence prediction. However, the model's performance is constrained by the quality of historical data, and prediction deviations caused by sudden risk events require additional repricing costs.

[0007] 4. Multi-Source Data Collaboration Stage: Building a Full-Scenario Intelligent Pricing System. To overcome limitations in data dimensions and scenario adaptability, the industry began exploring integrated pricing models that combine internal network data and public network information, driving the pricing system towards full-scenario intelligent upgrading. In 2020, the Ravizza team designed a multi-source data weighted pricing model, integrating enterprise internal network claims data with public network scenario data to expand the coverage of risk factors. However, the incomplete availability calibration rules for heterogeneous data led to low data fusion efficiency. In 2022, Brownlee et al. automatically extracted unstructured data features through large models and combined them with fuzzy membership functions to quantify risk uncertainty, providing theoretical support for dynamic pricing. However, the model's decision-making logic lacked transparency, resulting in high compliance interpretation costs. In 2023, Adacher proposed a large model-driven customer dynamic profiling system, optimizing pricing strategies through real-time data updates and improving personalized adaptability. However, its system architecture is complex, posing high deployment and maintenance barriers for small and medium-sized insurance companies. In 2024, Dabachine et al. combined large models with reinforcement learning, which improved the accuracy of risk prediction by 12.7% compared with traditional machine learning algorithms. However, rule conflicts and response delays still exist in the real-time price adjustment process, affecting the real-time performance and accuracy of pricing.

[0008] Therefore, we need to discuss this further. Summary of the Invention

[0009] The purpose of this application is to provide a short-term insurance dynamic pricing method based on an AI large model, and the specific technical solution is as follows:

[0010] A dynamic pricing method for short-term insurance based on an AI-powered large model includes: S1, constructing a multimodal risk dynamic perception architecture and outputting a unified dynamic risk vector of fixed length; S2, establishing a real-time risk pricing adaptive mechanism based on the unified dynamic risk vector of fixed length output in S1 to achieve real-time pricing decisions across insurance types; S3, designing an intelligent strategy generation and conflict resolution system to optimize the real-time pricing decisions across insurance types in S2, ensuring that pricing decisions achieve the best balance between market competitiveness and profitability while meeting actuarial constraints and regulatory requirements; S4, constructing a complete decision traceability mechanism using multimodal interpretable artificial intelligence technology, generating attribution reports based on the output data of S1-S3, and realizing the transformation of pricing decisions from a "black box" to a "white box".

[0011] The S1 module constructs a multimodal risk dynamic perception architecture, including: S1.1, a multimodal data access and preprocessing module, specifically comprising an unstructured data processing submodule and a structured data processing submodule; the unstructured data processing submodule processes medical text, vehicle damage images, and voice recordings, with text data undergoing automated cleaning, denoising, and standardization; image data undergoing normalization, compression, and data augmentation; and voice data converted to text using automatic speech recognition technology; the structured data processing submodule processes policy information, historical claims records, credit scores, and real-time sensor data, and performs data verification, cleaning, and statistical feature generation; S1.2, an AI large-scale model feature encoding module, specifically comprising a text feature encoder, a visual feature extractor, and a structured data embedder; the text feature encoder uses a large language model trained on insurance domain corpora to perform fine-grained entity recognition, sentiment analysis, and medical knowledge graph linking, extracting risk dimensions such as accident liability ambiguity, loss severity implications, and medical diagnosis correlations from the text data; the visual feature extractor is integrated into the vehicle damage and medical image processing submodules. The system trains a large visual model on a dataset, identifies damaged parts using an object detection network, quantifies the degree of damage using image segmentation and regression models, and generates a visual damage severity index and an imaging risk score. A structured data embedder uses embedding layers to map categorical features to a high-dimensional vector space, normalizes numerical features, and uses a time-series encoder to capture the changing patterns of time-series data such as driving behavior. S1.3: A dynamic risk feature fusion module is constructed, specifically including a cross-modal feature alignment submodule and a deep nonlinear fusion submodule. The cross-modal feature alignment submodule uses a cross-attention mechanism to achieve bidirectional information interaction between multimodal features. By calculating the correlation weights between different modal features, it achieves alignment and complementarity at the feature semantic level. When consistency between different modal features is detected, or when there are significant differences between text descriptions and image features, it automatically generates an inconsistent risk marker. The deep nonlinear fusion submodule contains a multi-layer fully connected neural network that performs nonlinear transformation and information compression on the aligned high-dimensional feature vector. It learns the complex interactions between different features through a multi-layer perceptron and finally outputs a unified dynamic risk vector of fixed length.

[0012] The S1 framework for constructing a multimodal risk dynamic perception architecture also includes: S1.4, which constructs a streaming update and real-time inference module, specifically including an event-driven processing submodule and an incremental inference submodule. The event-driven processing submodule monitors the data input port and automatically triggers the feature recalculation process when it receives new driving behavior data, image uploads, or voice recordings. The incremental inference module adopts a lightweight update strategy, which only runs the corresponding feature encoder for new data and quickly combines the new features with the cached historical fusion vectors through an incremental fusion network to generate an updated dynamic risk vector.

[0013] The S1.4 module for building streaming updates and real-time inference also includes a version management submodule, which records and tracks the feature vector update process to ensure the traceability of risk profile evolution.

[0014] The establishment of a real-time risk pricing adaptive mechanism in S2 includes: S2.1, constructing a multi-source streaming data acquisition module, which receives user behavior data from various insurance types in real time through a distributed message middleware: for auto insurance, it collects driving behavior sequences transmitted by OBD devices; for health insurance, it acquires time-series data of physiological indicators monitored by wearable devices; for accident insurance, it collects user movement trajectory and activity pattern data; and for life insurance, it receives data from regular physical examinations and lifestyle monitoring. After preprocessing, this data is input into a dynamic feature extraction engine based on a large language model. This engine uses a pre-trained language model to perform unified representation learning on multimodal time-series data, captures long-term dependencies through an attention mechanism, and uses online fine-tuning technology to make the model continuously adapt to changes in data distribution; S2.2, constructing a dynamic risk scoring model, using a large language model as the core architecture of sequence modeling, and utilizing powerful contextual understanding capabilities to process streaming data; specifically including: S 2.21. A Transformer-based temporal encoder is used to extract features and identify risk patterns from streaming data of various insurance types. 2.22. A cue learning technique is employed to transform the real-time risk scoring task into a text generation problem using a language model, achieving risk assessment through dynamically constructed cue templates. 2.23. A thought chain technique is used to enable the model to demonstrate the reasoning process from raw data to risk scoring. 2.24. Efficient parameter fine-tuning methods, such as LoRA, are applied to achieve rapid online adaptation to the basic large model. 2.3. A multi-task risk assessment framework based on a large language model is constructed, employing a memory-enhanced large model architecture to maintain a dynamically updated risk knowledge base, storing recent risk assessment patterns and information on sudden risk events. Simultaneously, the reasoning capabilities of the large model enable in-depth analysis of complex risk scenarios and assessment of cross-insurance risk contagion effects, ultimately achieving real-time pricing decisions across insurance types.

[0015] S2.2 also includes the following when building a dynamic risk scoring model: S2.25, designing an event-triggered model update mechanism, which automatically starts the incremental learning process when an abnormal pattern is detected, ensuring that the risk assessment model is updated within seconds.

[0016] The design of the intelligent strategy generation and conflict resolution system in S3 includes: S3.1, using a large language model fine-tuned with actuarial knowledge as the decision core to construct the strategy generation module. For four short-term insurance businesses—auto insurance, health insurance, accident insurance, and life insurance—a multi-head attention mechanism based on the Transformer architecture is constructed to handle the risk feature vectors and pricing factors specific to each insurance type. S3.2, using a multi-objective constraint optimization framework based on a large model to construct the conflict resolution module, specifically including: S3.21, utilizing the semantic understanding capability of the large language model to transform actuarial constraints into computable vector representations; S3.22, leveraging the contextual capabilities of the large model to parse and adapt constraints in real time. S3.23. Design a reward shaping mechanism based on a large model. By analyzing market feedback and profitability of historical pricing decisions, automatically adjust the reward function weights of pricing strategies for each insurance type. S3.24. To address strategy conflicts among multiple insurance types, adopt a course learning strategy guided by a large model. Through sequence-to-sequence modeling, gradually optimize the synergy of pricing strategies for each insurance type. S3.3. Construct a compliance verification module for a real-time regulatory compliance inspection system based on a large language model. Utilize the large model's deep understanding of regulatory documents and company policies. Through textual implication and semantic similarity calculation, automatically detect the consistency between the generated strategy and regulatory requirements. Specifically, this includes using the large model's few-shot learning capabilities to quickly adapt to regulatory policy updates, demonstrating the logical process of compliance judgment through thought chain technology, and generating interpretable compliance reports.

[0017] The compliance verification module in S3.3 also includes a multi-granularity feedback mechanism based on a large model, which provides real-time optimization suggestions throughout the entire process of strategy generation, conflict resolution, and compliance checks, ensuring that the pricing strategy achieves the best balance between market competitiveness and profitability while meeting actuarial constraints and regulatory requirements.

[0018] The complete decision tracing mechanism in S4 includes: S4.1, establishing a decision logic parsing module based on thinking chain technology, using a large language model fine-tuned with actuarial knowledge to semantically reconstruct the entire pricing decision process, specifically including: S4.11, using attention weight visualization technology to identify the key risk factors with the greatest impact on the final pricing from the input features; S4.12, utilizing the sequence-to-sequence generation capability of the large model to transform the complex numerical calculation process into a decision logic chain described in natural language; S4.13, constructing domain-specific interpretation templates through prompting engineering technology to ensure that the generated attribution report conforms to actuarial professional standards; S4.2, constructing a multi-granularity interpretable output architecture at the attribution report generation level, specifically including: S4.21, based on Transfor... Mer's attribution algorithm calculates the relative importance of each risk factor in pricing decisions through a hierarchical attention mechanism; S4.22, leveraging the text generation capabilities of the large model, it automatically generates a complete report containing risk factor weight distribution, decision logic path, and actuarial basis; S4.23, employing knowledge graph enhancement technology, it correlates pricing decisions with insurance terms, regulatory requirements, and actuarial principles, providing authoritative references for decision-making; S4.3, it constructs a multi-version report generation mechanism based on the large model to achieve the dual goals of regulatory compliance and customer communication. Through domain-adaptive fine-tuning technology, the large model can generate explanatory reports with corresponding levels of detail and expression according to different audiences: reports for regulatory agencies emphasize actuarial compliance and statistical significance, while reports for customers emphasize easy understanding.

[0019] When building a multi-version report generation mechanism based on a large model in S4.3, a real-time compliance check function is also included. The large model is used to verify the compliance of the generated explanations to ensure that all statements meet regulatory requirements.

[0020] The beneficial effects of this application lie in its risk dynamic perception architecture built upon a multimodal large-scale model, achieving a leap from "shallow statistics" to "deep semantic understanding." Through deep analysis of unstructured text using a large language model and intelligent extraction of image features using a large visual model, the system can capture deep risk patterns that traditional methods cannot identify. For example, in the auto insurance field, the system can not only analyze driving behavior data but also automatically assess the degree of damage from accident photos; in the health insurance field, it can process medical examination report data and identify potential health risks from medical texts. The real-time risk pricing mechanism built upon the online learning capabilities of the large model achieves a technological leap from "periodic retraining" to "second-level risk updates." Through a streaming processing architecture and incremental learning technology, the system can complete risk assessment updates within milliseconds after receiving new driving behavior data, health monitoring indicators, and other real-time information. Compared to traditional pricing models updated quarterly or monthly, this system significantly improves the response speed to sudden risk events, especially in the accident and auto insurance fields, enabling real-time responses to environmental risk factors such as sudden weather changes and road conditions. This intelligent strategy generation system, built upon large language models and reinforcement learning, effectively solves the challenge of balancing market competition, profit targets, and regulatory requirements. Through a multi-agent reinforcement learning framework, the system generates optimal pricing strategies for different insurance products, while simultaneously embedding compliance checks in real-time during the decision-making process. The interpretable pricing scheme, built using the natural language generation capabilities of large models, completely transforms the "black box" decision-making dilemma of traditional pricing models. The system automatically generates attribution reports for each pricing decision, clearly demonstrating the impact weights of key risk factors and the decision-making logic path, improving customer acceptance and reducing complaint rates. This end-to-end transparency not only meets increasingly stringent regulatory requirements but also establishes customer relationships based on understanding and trust, laying a solid foundation for the long-term development of the insurance business. Through microservice architecture and model parallel technology, the system maintains millisecond-level response times while supporting parallel processing of multiple insurance products. In actual deployment, the system successfully supports over one million pricing requests per day, spanning four major business areas: auto insurance, accident insurance, health insurance, and life insurance. This high scalability ensures that the system can adapt to the rapid growth of the insurance business and continuous product innovation. In summary, this patented technology, through the innovative application of AI big data models, has achieved comprehensive breakthroughs in risk identification accuracy, pricing response speed, multi-objective decision optimization, business transparency, and system efficiency, providing a new generation of intelligent pricing solutions for short-term insurance business. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the application process. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to specific embodiments and accompanying drawings. It should be understood that these descriptions are merely exemplary and not intended to limit the scope of this application. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concepts of this application.

[0023] A short-term insurance dynamic pricing method based on an AI large model includes:

[0024] S1. Construct a multimodal risk dynamic perception architecture and output a unified dynamic risk vector of fixed length. Constructing the multimodal risk dynamic perception architecture includes:

[0025] S1.1 Construct a multimodal data access and preprocessing module, specifically including an unstructured data processing submodule and a structured data processing submodule. The unstructured data processing submodule is used to process medical text, vehicle damage images, and voice recordings. Text data undergoes automated cleaning, noise reduction, and standardization. Image data undergoes normalization, compression, and data enhancement. Voice data is converted into text using automatic speech recognition technology. The structured data processing submodule is used to process policy information, historical claims records, credit scores, and real-time sensor data, and performs data verification, cleaning, and statistical feature generation.

[0026] S1.2 Constructing an AI large-scale model feature encoding module, specifically including a text feature encoder, a visual feature extractor, and a structured data embedder; among them, the text feature encoder uses a large language model trained on insurance domain corpus to perform fine-grained entity recognition, sentiment analysis, and medical knowledge graph linking, extracting risk dimensions such as ambiguity of accident liability, indication of the degree of loss, and correlation with medical diagnosis from text data; the visual feature extractor is integrated into a large visual model trained on vehicle damage and medical image datasets, identifies damaged parts through object detection networks, quantifies the degree of damage using image segmentation and regression models, and generates a visual damage severity index and imaging risk score; the structured data embedder uses an embedding layer to map categorical features to a high-dimensional vector space, normalizes numerical features, and uses a temporal encoder to capture the change patterns of temporal data such as driving behavior.

[0027] S1.3 Construct a dynamic risk feature fusion module, specifically including a cross-modal feature alignment submodule and a deep nonlinear fusion submodule. The cross-modal feature alignment submodule uses a cross-attention mechanism to achieve bidirectional information interaction between multimodal features. By calculating the correlation weights between different modal features, it achieves alignment and complementarity at the feature semantic level. When consistency between different modal features is detected, or when there are significant differences between text descriptions and image features, it automatically generates an inconsistent risk label. The deep nonlinear fusion submodule contains a multilayer fully connected neural network. It performs nonlinear transformation and information compression on the aligned high-dimensional feature vector. Through a multilayer perceptron, it learns the complex interactions between different features and finally outputs a unified dynamic risk vector of fixed length.

[0028] S1.4 Construct a streaming update and real-time inference module, specifically including an event-driven processing submodule and an incremental inference submodule. The event-driven processing submodule monitors the data input port and automatically triggers the feature recalculation process when new driving behavior data, image uploads, or voice recordings are received. The incremental inference module adopts a lightweight update strategy, running only the corresponding feature encoder for new data. It then uses an incremental fusion network to quickly combine the new features with cached historical fusion vectors to generate an updated dynamic risk vector. The streaming update and real-time inference module also includes a version management submodule to record and track the feature vector update process, ensuring the traceability of the risk profile's evolution.

[0029] S2. Based on the fixed-length unified dynamic risk vector output in S1, establish a real-time risk pricing adaptive mechanism to achieve real-time pricing decisions across insurance types. Establishing the real-time risk pricing adaptive mechanism includes:

[0030] S2.1 Construct a multi-source streaming data acquisition module to receive user behavior data from various insurance types in real time through a distributed message middleware: for auto insurance, it collects driving behavior sequences transmitted by OBD devices; for health insurance, it acquires time-series data of physiological indicators monitored by wearable devices; for accident insurance, it collects user movement trajectory and activity pattern data; and for life insurance, it receives data from regular physical examinations and lifestyle monitoring. After preprocessing, this data is input into a dynamic feature extraction engine based on a large language model. This engine uses a pre-trained language model to perform unified representation learning on multimodal time-series data, captures long-term dependencies through an attention mechanism, and uses online fine-tuning technology to make the model continuously adapt to changes in data distribution.

[0031] S2.2 Construct a dynamic risk scoring model, employing a large language model as the core architecture for sequence modeling, and leveraging powerful contextual understanding capabilities to process streaming data. Specifically, this includes: S2.21, a Transformer-based temporal encoder for feature extraction and risk pattern recognition of streaming data across various insurance types. S2.22, Employing cue learning technology to transform the real-time risk scoring task into a text generation problem using a language model, achieving risk assessment through dynamically constructed cue templates. S2.23, Utilizing thought chain technology to enable the model to demonstrate the reasoning process from raw data to risk scoring. S2.24, Applying efficient parameter fine-tuning methods, such as LoRA, to achieve rapid online adaptation to the basic large model. S2.25, Designing an event-triggered model update mechanism that automatically initiates an incremental learning process when an abnormal pattern is detected, ensuring the risk assessment model is updated within seconds.

[0032] S2.3 Construct a multi-task risk assessment framework based on a large language model. Adopt a memory-enhanced large model architecture to maintain a dynamically updated risk knowledge base, storing recent risk assessment patterns and information on sudden risk events. At the same time, through the reasoning capabilities of the large model, achieve in-depth analysis of complex risk scenarios and assessment of cross-insurance risk contagion effects. Ultimately, achieve real-time pricing decisions across insurance types.

[0033] S3. Design an intelligent strategy generation and conflict resolution system to optimize the real-time pricing decisions across insurance types in S2, ensuring that pricing decisions achieve the best balance between market competitiveness and profitability while meeting actuarial constraints and regulatory requirements. The design of the intelligent strategy generation and conflict resolution system includes:

[0034] S3.1. A large language model fine-tuned with actuarial knowledge is used as the core of the decision-making process to build a strategy generation module. For four types of short-term insurance businesses, namely auto insurance, health insurance, accident insurance and life insurance, a multi-head attention mechanism based on the Transformer architecture is constructed to handle the risk feature vectors and pricing factors unique to each type of insurance.

[0035] S3.2. A conflict resolution module is constructed using a multi-objective constraint optimization framework based on a large model. Specifically, this includes: S3.21. Utilizing the semantic understanding capabilities of the large language model, actuarial constraints are transformed into computable vector representations. S3.22. Through the contextual capabilities of the large model, constraints are parsed and adapted to changes in real time. S3.23. A reward shaping mechanism based on the large model is designed to automatically adjust the reward function weights of pricing strategies for each insurance product by analyzing market feedback and profitability of historical pricing decisions. S3.24. To address strategy conflicts among multiple insurance products, a large model-guided learning strategy is adopted, using a sequence-to-sequence modeling approach to gradually optimize the synergy of pricing strategies for each insurance product.

[0036] S3.3. A real-time regulatory compliance inspection system based on a large language model constructs a compliance verification module. Leveraging the large model's deep understanding of regulatory documents and company policies, it automatically detects the consistency between generated strategies and regulatory requirements through textual entailment and semantic similarity calculations. Specifically, it utilizes the large model's few-shot learning capabilities to quickly adapt to regulatory policy updates, demonstrates the logical process of compliance judgment through thought chain technology, and generates interpretable compliance reports. The compliance verification module also includes a multi-granularity feedback mechanism based on the large model, providing real-time optimization suggestions throughout the entire process of strategy generation, conflict resolution, and compliance inspection. This ensures that pricing strategies achieve the optimal balance between market competitiveness and profitability while meeting actuarial constraints and regulatory requirements.

[0037] S4. Employ multimodal interpretable artificial intelligence technology to construct a complete decision-making traceability mechanism. Based on the output data of S1-S3, generate attribution reports to transform pricing decisions from a "black box" to a "white box." Constructing a complete decision-making traceability mechanism includes:

[0038] S4.1 Establish a decision logic analysis module based on thinking chain technology. Utilize a large language model fine-tuned with actuarial knowledge to semantically reconstruct the entire pricing decision-making process. Specifically, this includes: S4.11 Employing attention weight visualization technology to identify the key risk factors among input features that have the greatest impact on the final pricing. S4.12 Leveraging the sequence-to-sequence generation capability of the large model, transforming complex numerical calculation processes into a decision logic chain described in natural language. S4.13 Constructing domain-specific explanation templates through prompting engineering techniques to ensure that the generated attribution report conforms to actuarial professional standards.

[0039] S4.2 Construct a multi-granularity interpretable output architecture for attribution report generation, specifically including: S4.21, a Transformer-based attribution algorithm that calculates the relative importance of each risk factor in pricing decisions through a hierarchical attention mechanism. S4.22, Utilizing the text generation capabilities of large models to automatically generate a complete report containing risk factor weight distribution, decision logic path, and actuarial basis. S4.23, Employing knowledge graph enhancement technology to correlate pricing decisions with insurance terms, regulatory requirements, and actuarial principles, providing authoritative references to decision-making.

[0040] S4.3 Construct a multi-version report generation mechanism based on a large model to achieve the dual goals of regulatory compliance and client communication. Through domain-adaptive fine-tuning technology, the large model can generate explanatory reports with varying levels of detail and expression depending on the audience: reports for regulatory agencies emphasize actuarial compliance and statistical significance, while reports for clients emphasize clarity and accessibility. The multi-version report generation mechanism also includes a real-time compliance check function, using the large model to verify the compliance of the generated explanations, ensuring that all statements meet regulatory requirements.

[0041] In its implementation, the dynamic risk assessment module uses a fine-tuned LLaMA model as its infrastructure, employing LoRA technology for efficient parameter fine-tuning to imbue it with insurance domain knowledge. This module receives multimodal input data: for auto insurance, it analyzes driving behavior sequences and vehicle conditions in real time; for health insurance, it processes physiological indicators monitored by wearable devices; and for accident insurance, it assesses user activity trajectories and environmental risks. The system incorporates an online learning mechanism, using incremental learning algorithms to update model parameters minute-by-minute, automatically triggering model recalibration when a sudden risk event is detected. The pricing decision module employs a multi-agent reinforcement learning framework, designing dedicated reward functions for different insurance types, while a compliance verification module ensures all pricing decisions comply with regulatory requirements.

Claims

1. A short-term insurance dynamic pricing method based on an AI large-scale model, characterized in that, include: S1. Construct a multimodal risk dynamic perception architecture and output a unified dynamic risk vector of fixed length; S2. Based on the fixed-length unified dynamic risk vector output in S1, establish a real-time risk pricing adaptive mechanism to realize real-time pricing decisions across insurance types. S3. Design an intelligent strategy generation and conflict resolution system to optimize the real-time pricing decisions across insurance types in S2, ensuring that the pricing decisions achieve the best balance between market competitiveness and profit targets while meeting actuarial constraints and regulatory requirements. S4. A complete decision-making traceability mechanism is constructed using multimodal interpretable artificial intelligence technology. Based on the output data of S1-S3, an attribution report is generated to realize the transformation of pricing decisions from a "black box" to a "white box".

2. The short-term insurance dynamic pricing method based on an AI large model as described in claim 1, characterized in that, The construction of the multimodal risk dynamic perception architecture in S1 includes: S1.1 Construct a multimodal data access and preprocessing module, specifically including an unstructured data processing submodule and a structured data processing submodule. The unstructured data processing submodule processes medical text, vehicle damage images, and voice recordings. Text data undergoes automated cleaning, noise reduction, and standardization; image data undergoes normalization, compression, and data augmentation; and voice data is converted into text using automatic speech recognition technology. The structured data processing submodule processes policy information, historical claims records, credit scores, and real-time sensor data, and performs data verification, cleaning, and statistical feature generation. S1.2 Constructing an AI large-scale model feature encoding module, specifically including a text feature encoder, a visual feature extractor, and a structured data embedder; among them, the text feature encoder uses a large language model trained on insurance domain corpus to perform fine-grained entity recognition, sentiment analysis, and medical knowledge graph linking, extracting risk dimensions such as ambiguity of accident liability, indication of the degree of loss, and correlation with medical diagnosis from text data; the visual feature extractor is integrated into a large visual model trained on vehicle damage and medical image datasets, identifies damaged parts through an object detection network, quantifies the degree of damage using image segmentation and regression models, and generates a visual damage severity index and imaging risk score; the structured data embedder uses an embedding layer to map categorical features to a high-dimensional vector space, normalizes numerical features, and uses a temporal encoder to capture the changing patterns of temporal data such as driving behavior; S1.3 Construct a dynamic risk feature fusion module, specifically including a cross-modal feature alignment submodule and a deep nonlinear fusion submodule. The cross-modal feature alignment submodule uses a cross-attention mechanism to achieve bidirectional information interaction between multimodal features. By calculating the correlation weights between different modal features, it achieves alignment and complementarity at the feature semantic level. When consistency between different modal features is detected, or when there are significant differences between text descriptions and image features, it automatically generates an inconsistent risk label. The deep nonlinear fusion submodule contains a multilayer fully connected neural network. It performs nonlinear transformation and information compression on the aligned high-dimensional feature vector. Through a multilayer perceptron, it learns the complex interactions between different features and finally outputs a unified dynamic risk vector of fixed length.

3. The short-term insurance dynamic pricing method based on an AI large model as described in claim 2, characterized in that, The construction of the multimodal risk dynamic perception architecture in S1 also includes: S1.4 Construct a streaming update and real-time inference module, specifically including an event-driven processing submodule and an incremental inference submodule. The event-driven processing submodule monitors the data input port and automatically triggers the feature recalculation process when it receives new driving behavior data, image uploads, or voice recordings. The incremental inference module adopts a lightweight update strategy, which only runs the corresponding feature encoder for new data and quickly combines the new features with the cached historical fusion vectors through an incremental fusion network to generate an updated dynamic risk vector.

4. The short-term insurance dynamic pricing method based on an AI large model as described in claim 3, characterized in that, The S1.4 module for building streaming update and real-time inference also includes a version management submodule, which records and tracks the feature vector update process to ensure the traceability of risk profile evolution.

5. The short-term insurance dynamic pricing method based on an AI large model as described in claim 2, characterized in that, The establishment of the real-time risk pricing adaptive mechanism in S2 includes: S2.1 Construct a multi-source streaming data acquisition module to receive user behavior data from various insurance types in real time through a distributed message middleware: in the auto insurance field, it collects driving behavior sequences transmitted by OBD devices; in the health insurance field, it obtains time-series data of physiological indicators monitored by wearable devices; in the accident insurance field, it collects user movement trajectory and activity pattern data; and in the life insurance field, it receives regular physical examination and lifestyle monitoring data. After preprocessing, these data are input into a dynamic feature extraction engine based on a large language model. This engine uses a pre-trained language model to perform unified representation learning on multimodal time-series data, captures long-term dependencies through an attention mechanism, and uses online fine-tuning technology to make the model continuously adapt to changes in data distribution. S2.2 Construct a dynamic risk scoring model, using a large language model as the core architecture for sequence modeling, and leveraging powerful contextual understanding capabilities to process streaming data; specifically including: S2.

21. A Transformer-based temporal encoder is used to extract features and identify risk patterns in streaming data of various types of insurance. S2.

22. Employ prompt learning technology to transform the real-time risk scoring task into a text generation problem of a language model, and achieve risk assessment by dynamically constructing prompt templates; S2.

23. Utilize the thinking chain technique to enable the model to demonstrate the reasoning process from raw data to risk score; S2.

24. Apply efficient parameter fine-tuning methods, such as LoRA, to achieve rapid online adaptation to the basic large model; S2.3 Construct a multi-task risk assessment framework based on a large language model. Adopt a memory-enhanced large model architecture to maintain a dynamically updated risk knowledge base, storing recent risk assessment patterns and information on sudden risk events. At the same time, through the reasoning capabilities of the large model, achieve in-depth analysis of complex risk scenarios and assessment of cross-insurance risk contagion effects. Ultimately, achieve real-time pricing decisions across insurance types.

6. The short-term insurance dynamic pricing method based on an AI large model as described in claim 2, characterized in that, The construction of the dynamic risk scoring model in S2.2 also includes: S2.

25. Design an event-triggered model update mechanism that automatically starts the incremental learning process when an abnormal pattern is detected, ensuring that the risk assessment model is updated within seconds.

7. The short-term insurance dynamic pricing method based on an AI large model as described in claim 5, characterized in that, The design of the intelligent strategy generation and conflict resolution system in S3 includes: S3.

1. A large language model fine-tuned with actuarial knowledge is used as the core decision-making module to build a strategy generation module. For four types of short-term insurance businesses, namely auto insurance, health insurance, accident insurance and life insurance, a multi-head attention mechanism based on the Transformer architecture is built to handle the risk feature vectors and pricing factors unique to each type of insurance. S3.

2. A conflict resolution module is constructed using a multi-objective constraint optimization framework based on a large model, specifically including: S3.

21. Utilize the semantic understanding capabilities of large language models to transform actuarial constraints into computable vector representations; S3.

22. By leveraging the contextual capabilities of large models, real-time analysis and adaptation to changes in constraints are achieved. S3.

23. Design a reward shaping mechanism based on a large model, and automatically adjust the reward function weights of the pricing strategies for each type of insurance by analyzing the market feedback and profitability of historical pricing decisions. S3.

24. To address the strategic conflicts among multiple insurance products, a large-model-guided course learning strategy is adopted. Through a sequence-to-sequence modeling approach, the pricing strategy coordination of each insurance product is gradually optimized. S3.

3. The real-time regulatory compliance inspection system based on a large language model constructs a compliance verification module. It utilizes the large model's deep understanding of regulatory documents and company policies, and automatically detects the consistency between the generated strategy and regulatory requirements through textual implication and semantic similarity calculation. Specifically, it uses the large model's few-shot learning capability to quickly adapt to regulatory policy updates, demonstrates the logical process of compliance judgment through thought chain technology, and generates interpretable compliance reports.

8. The short-term insurance dynamic pricing method based on an AI large model as described in claim 7, characterized in that, The compliance verification module in S3.3 also includes a multi-granularity feedback mechanism based on a large model, which provides real-time optimization suggestions throughout the entire process of strategy generation, conflict resolution, and compliance checks, ensuring that the pricing strategy achieves the best balance between market competitiveness and profitability while meeting actuarial constraints and regulatory requirements.

9. The short-term insurance dynamic pricing method based on an AI large model as described in claim 7, characterized in that, The complete decision tracing mechanism constructed in S4 includes: S4.1 Establish a decision logic analysis module based on thinking chain technology, and use a large language model finely tuned with actuarial knowledge to semantically reconstruct the entire pricing decision process, specifically including: S4.

11. Employ attention weight visualization technology to identify the key risk factors among the input features that have the greatest impact on the final pricing; S4.

12. Utilize the sequence-to-sequence generation capability of large models to transform complex numerical calculation processes into decision logic chains described in natural language. S4.

13. By prompting engineering technology-specific interpretation templates, ensure that the generated attribution reports comply with actuarial professional standards; S4.2 Construct a multi-granularity interpretable output architecture for attribution report generation, specifically including: S4.

21. A Transformer-based attribution algorithm calculates the relative importance of each risk factor in pricing decisions through a hierarchical attention mechanism. S4.

22. Utilize the text generation capabilities of large models to automatically generate a complete report that includes risk factor weight distribution, decision logic path, and actuarial basis; S4.

23. Employ knowledge graph enhancement technology to correlate pricing decisions with insurance terms, regulatory requirements, and actuarial principles, providing authoritative decision-making basis references; S4.3 Build a multi-version report generation mechanism based on a large model to achieve the dual goals of regulatory compliance and customer communication. Through domain-adaptive fine-tuning technology, the large model can generate explanatory reports with corresponding levels of detail and expression according to different audiences: reports for regulatory agencies emphasize actuarial compliance and statistical significance, while reports for customers emphasize easy understanding.

10. The short-term insurance dynamic pricing method based on an AI large model as described in claim 9, characterized in that, The S4.3 mechanism for building a multi-version report generation mechanism based on a large model also includes a real-time compliance check function, which uses the large model to verify the compliance of the generated explanations and ensures that all statements meet regulatory requirements.

Images