Annuity asset data processing method and device based on large model

Abstract

This invention discloses a method and apparatus for processing annuity asset data based on a large model. The method includes: collecting multi-source heterogeneous data from annuity business scenarios and processing it to obtain standardized original basic annuity feature data; extracting annuity operational feature data and risk feature data from the original basic feature data, and inputting them into a pre-trained annuity operational feature data recognition model and an annuity risk assessment model to obtain corresponding annuity operational feature data and quantitative risk assessment results; inputting the original basic feature data and the output results of the two models into a configuration optimization model, and using reinforcement learning for multi-round iterative optimization to adaptively generate annuity asset allocation ratio data. This invention enables multi-dimensional dynamic analysis of annuity asset data, deeply mining the potential features of the original data, improving data information utilization, and effectively improving the accuracy of annuity data analysis and adaptability to complex business scenarios through large model correlation analysis and reinforcement learning dynamic optimization.

Description

Technical Field

[0001] This invention relates to the field of big data processing technology, and in particular to a method and apparatus for processing annuity asset data based on a large model. Background Technology

[0002] This section is intended to provide background or context for the embodiments of the invention set forth in the claims. The description herein is not an admission that it is prior art simply because it is included in this section.

[0003] Current pension asset data processing generally employs fixed mathematical models and single intelligent models for data computation and feature analysis, resulting in a static and rigid processing mode. Traditional methods rely on fixed formulas and manually preset parameters, leading to fixed data analysis dimensions, rigid and closed processing logic, and a lack of the ability to perform correlational reasoning and judgment using dedicated large-scale models. This makes it impossible to conduct holistic linkage analysis and multi-dimensional feature collaborative analysis of multiple types of related pension asset data. It can only perform shallow, single-dimensional static feature mining of asset data, failing to fully uncover hidden correlations between multi-source data. The utilization rate of effective features in the original data is extremely low, easily leading to one-sided feature extraction and biased data analysis. Furthermore, traditional technologies lack dynamic iterative optimization mechanisms, resulting in a one-time fixed computation output. This prevents continuous optimization of computation results by combining multi-dimensional data features, resulting in a single and rigid data judgment output. Ultimately, this leads to limited accuracy in subsequent asset analysis results and insufficient adaptability to complex scenarios, failing to meet the needs of refined and dynamic pension asset big data processing technologies. Summary of the Invention

[0004] This invention provides a method for processing annuity asset data based on a large model, addressing the technical problems of existing annuity asset data processing models being static and rigid, having a single analytical dimension, shallow feature mining of raw data, low information utilization, and lacking large model correlation analysis capabilities and dynamic iterative optimization mechanisms, resulting in low data analysis accuracy and poor scenario adaptability. The method includes: Collect multi-source heterogeneous data in the context of annuity business, preprocess the multi-source heterogeneous data, and obtain standardized annuity raw basic feature data; Annuity operational characteristic data is extracted from the original basic annuity characteristic data, and the annuity operational characteristic data is input into the annuity operational characteristic data recognition model to obtain the annuity operational characteristic data recognition result; annuity risk characteristic data is extracted from the original basic annuity characteristic data, and the annuity risk characteristic data is input into the annuity risk assessment model to obtain the annuity quantitative risk assessment result; the annuity operational characteristic data recognition model is pre-trained and generated from historical annuity operational characteristic data and corresponding recognition result samples, and the annuity risk assessment model is pre-trained and generated from historical annuity risk characteristic data and corresponding quantitative risk assessment result samples. The original basic characteristic data of annuities, the identification results of operational characteristic data of annuities, and the quantitative risk assessment results of annuities are input into the configuration optimization model. The configuration optimization model is based on reinforcement learning and performs multiple rounds of iterative optimization to adaptively generate annuity asset allocation ratio data.

[0005] This invention also provides a large-model-based annuity asset data processing device to address the technical problems of existing annuity asset data processing models being static and fixed, having a single analytical dimension, shallow feature mining of raw data, low information utilization, and lacking large-model correlation analysis capabilities and dynamic iterative optimization mechanisms, resulting in low data analysis accuracy and poor scenario adaptability. The device includes: The data acquisition unit is used to collect multi-source heterogeneous data in the annuity business scenario, and to preprocess the multi-source heterogeneous data to obtain standardized annuity raw basic feature data. The parallel parsing unit is used to extract annuity operational feature data from the original basic annuity feature data, input the annuity operational feature data into the annuity operational feature data recognition model, and obtain annuity operational feature data recognition results; it also extracts annuity risk feature data from the original basic annuity feature data, inputs the annuity risk feature data into the annuity risk assessment model, and obtains annuity quantitative risk assessment results; the annuity operational feature data recognition model is pre-trained and generated from historical annuity operational feature data and corresponding recognition result samples, and the annuity risk assessment model is pre-trained and generated from historical annuity risk feature data and corresponding quantitative risk assessment result samples. The processing unit is used to input the original basic characteristic data of annuities, the identification results of the operational characteristic data of annuities, and the quantitative risk assessment results of annuities into the configuration optimization model. The configuration optimization model is based on reinforcement learning to perform multiple rounds of iterative optimization and adaptively generate annuity asset allocation ratio data.

[0006] This invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the above-described method for processing annuity asset data based on a large model.

[0007] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for processing annuity asset data based on a large model.

[0008] This invention also provides a computer program product, which includes a computer program that, when executed by a processor, implements the above-described method for processing annuity asset data based on a large model.

[0009] In this embodiment of the invention, the beneficial technical effects of the annuity asset data processing scheme based on a large model are: This invention, relying on standardized annuity raw data, innovatively employs a dual-branch parallel parsing architecture. It simultaneously performs two independent large-scale model operations on the raw data: one path extracts annuity operational features and inputs them into a pre-trained annuity operational feature data recognition model to achieve quantitative identification of annuity operations; the other path extracts annuity risk features and inputs them into a pre-trained annuity risk assessment model to achieve quantitative assessment and analysis of asset risk. This invention differs from the traditional single-path, serial, shallow mining static processing mode. By using a dual-model parallel independent parsing approach, it deeply mines hidden correlation features between multi-source data, significantly improving the information utilization rate and feature mining completeness of the raw data. This effectively solves the problems of one-sided feature mining, information loss, and analytical bias in traditional techniques. Furthermore, relying on the correlation reasoning capabilities of the aforementioned annuity operational feature data recognition model and annuity risk assessment model, this invention performs holistic linked analysis of multiple types of related asset data, overcoming the limitations of traditional static, isolated, and single-dimensional data processing. Building upon this foundation, this invention further integrates three types of multidimensional data: original basic annuity characteristic data, annuity operational characteristic data identification results, and quantitative risk assessment results. These data are input into a configuration optimization model and, relying on reinforcement learning algorithms, achieve multi-round iterative optimization dynamic calculations. This continuously and adaptively optimizes the data processing logic and calculation results, completely breaking through the outdated processing mode of traditional static solidification and one-time output. It significantly improves the calculation accuracy, multidimensional correlation analysis capabilities, and adaptability to complex market scenarios in annuity asset data analysis, providing highly accurate and stable data support for the subsequent refined and intelligent allocation and output of annuity assets. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art 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. In the drawings: Figure 1 This is a flowchart illustrating a method for processing annuity asset data based on a large model, according to one embodiment of the present invention.

[0011] Figure 2 This is a schematic diagram illustrating the principle of annuity asset data processing based on a large model in one embodiment of the present invention.

[0012] Figure 3 This is a schematic diagram of the process for visualizing the output of decision explanation text in one embodiment of the present invention.

[0013] Figure 4 This is a schematic diagram of the structure of a pension asset data processing device based on a large model in one embodiment of the present invention.

[0014] Figure 5 This is a schematic diagram of a computer device structure according to an embodiment of the present invention. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. Here, the illustrative embodiments of the present invention and their descriptions are used to explain the present invention, but are not intended to limit the present invention.

[0016] The acquisition, storage, use, and processing of data in this application comply with relevant laws and regulations.

[0017] It should be noted that this application is specifically designed for the intelligent data processing of specialized annuity assets, adapting to a unique scenario. This adaptation fully aligns with the core business attributes and application characteristics of annuity assets, emphasizing long-term sustainability, stable operation, and dynamic assessment. Unlike traditional annuity processing technologies that rely solely on fixed mathematical models and single-dimensional data for static calculation and analysis, this application focuses on the full-process data characteristics of annuity asset management, covering heterogeneous data such as account data, asset management data, industry information, and operating financial reports. It constructs a multi-data foundation adapted to annuity-specific business, offering scenario adaptation advantages not found in traditional general financial data processing models. Addressing the complex data dimensions, dynamic changes in asset risk, and real-time fluctuations in business conditions inherent in annuity business, and the difficulty of achieving multi-dimensional data collaborative analysis and dynamic iterative optimization using traditional technologies, this application deeply adapts to the business needs of multi-source heterogeneous data linkage processing in annuities. It establishes a complete technical chain encompassing standardized data processing, parallel analysis of dual models, and intelligent iterative assessment. By combining multi-dimensional annuity business data with a reinforcement learning dynamic optimization mechanism, it achieves accurate risk quantification and intelligent data assessment and derivation. This technology effectively addresses the industry pain points of traditional annuity data processing techniques, such as rigidity, limited dimensions, insufficient feature mining, and weak dynamic adaptability. It precisely matches the refined, dynamic, and intelligent data processing needs of annuity assets, demonstrating outstanding scenario originality and unique value for financial asset management applications. Specifically, the uniqueness of the scenario this application applies to is reflected in the following aspects: First, the data dimensions are unique. Unlike traditional annuities that rely solely on standardized financial data, this scenario integrates structured asset management data with unstructured industry information and financial statement data to construct a comprehensive, multi-dimensional data foundation specific to annuities. Second, the processing mechanism is unique. It breaks away from the rigid model of static, single-operation, and serial processing in traditional annuities, achieving parallel analysis of business situation identification and risk assessment, resulting in more comprehensive data mining. Third, the adaptation requirements are unique. It can adapt to the long-term operation and dynamic fluctuations of annuity assets, achieving dynamic updates of data analysis results through iterative optimization, overcoming the technical limitations of traditional general-purpose technologies that cannot adapt to complex dynamic scenarios in annuity asset management.

[0018] Figure 1 This is a flowchart illustrating a method for processing annuity asset data based on a large model, as described in one embodiment of the present invention. Figure 1 As shown, the method includes the following steps: Step 101: Collect multi-source heterogeneous data in the annuity business scenario, preprocess the multi-source heterogeneous data, and obtain standardized annuity original basic feature data; Step 102: Extract annuity operational feature data from the original basic annuity feature data, input the annuity operational feature data into the annuity operational feature data recognition model, and obtain the annuity operational feature data recognition result; extract annuity risk feature data from the original basic annuity feature data, input the annuity risk feature data into the annuity risk assessment model, and obtain the annuity quantitative risk assessment result; the annuity operational feature data recognition model is pre-trained and generated from historical annuity operational feature data and corresponding recognition result samples, and the annuity risk assessment model is pre-trained and generated from historical annuity risk feature data and corresponding quantitative risk assessment result samples; Step 103: Input the original basic characteristic data of annuities, the identification results of the operational characteristic data of annuities, and the quantitative risk assessment results of annuities into the configuration optimization model. The configuration optimization model is based on reinforcement learning and performs multiple rounds of iterative optimization to adaptively generate annuity asset allocation ratio data.

[0019] The annuity asset data processing method based on a large model provided in this invention operates as follows: First, it comprehensively collects various multi-source heterogeneous data from annuity business scenarios, covering structured financial business data and unstructured industry text data. Through standardized preprocessing, data cleaning, feature normalization, and semantic vectorization, invalid and redundant information is eliminated, and data formats and dimensions are unified to generate high-quality, standardized original basic annuity feature data suitable for model computation, laying a precise data foundation for subsequent intelligent analysis. Second, based on the preprocessed original basic annuity feature data, it conducts parallel intelligent analysis with two branches, extracting corresponding annuity business situation feature data and annuity risk feature data. These two types of feature data are respectively input into a large-scale annuity business situation feature recognition model and an annuity risk assessment model, both trained with historical sample data. Relying on the strong correlation reasoning capabilities of the large model, it completes the quantitative recognition of situation features and the quantitative assessment of asset risks in parallel, simultaneously outputting accurate business situation recognition results and quantitative risk assessment results, achieving multi-dimensional, in-depth, and interconnected feature mining and analysis of annuity asset data. Finally, the three types of multi-dimensional core data—original basic annuity feature data, business situation identification results, and quantitative risk assessment results—are integrated and uniformly input into the iterative optimization model. Relying on reinforcement learning algorithms, multiple rounds of dynamic iterative optimization calculations are carried out to continuously and adaptively correct the data processing logic and analysis results. The model then dynamically outputs intelligent data judgment results that adapt to the dynamic changes in annuity business, thus completing a refined, dynamic, and intelligent processing flow for the entire annuity asset data.

[0020] The beneficial technical effects of the annuity asset data processing method based on a large model in this embodiment of the invention are: This invention, relying on standardized annuity raw data, innovatively employs a dual-branch parallel parsing architecture. It simultaneously performs two independent large-scale model operations on the raw data: one path extracts annuity operational features and inputs them into a pre-trained annuity operational feature data recognition model to achieve quantitative identification of annuity operations; the other path extracts annuity risk features and inputs them into a pre-trained annuity risk assessment model to achieve quantitative assessment and analysis of asset risk. This invention differs from the traditional single-path, serial, shallow mining static processing mode. By using a dual-model parallel independent parsing approach, it deeply mines hidden correlation features between multi-source data, significantly improving the information utilization rate and feature mining completeness of the raw data. This effectively solves the problems of one-sided feature mining, information loss, and analytical bias in traditional techniques. Furthermore, relying on the correlation reasoning capabilities of the aforementioned annuity operational feature data recognition model and annuity risk assessment model, this invention performs holistic linked analysis of multiple types of related asset data, overcoming the limitations of traditional static, isolated, and single-dimensional data processing. Building upon this foundation, this invention further integrates three types of multidimensional data: original basic annuity characteristic data, annuity operational characteristic data identification results, and quantitative risk assessment results. These data are input into a configuration optimization model and, relying on reinforcement learning algorithms, achieve multi-round iterative optimization dynamic calculations. This continuously and adaptively optimizes the data processing logic and calculation results, completely breaking through the outdated processing mode of traditional static solidification and one-time output. It significantly improves the calculation accuracy, multidimensional correlation analysis capabilities, and adaptability to complex market scenarios in annuity asset data analysis, providing highly accurate and stable data support for the subsequent refined and intelligent allocation and output of annuity assets.

[0021] The following is combined with Figures 2 to 3 This paper provides a detailed introduction to the data processing method for annuity assets based on a large model.

[0022] Figure 2 This is a schematic diagram illustrating the principle of annuity asset data processing based on a large model in one embodiment of the present invention. The present invention constructs a complete, holistic intelligent data processing architecture for annuity assets, encompassing multi-source data preprocessing, parallel intelligent parsing of dual models, multi-dimensional data fusion optimization, and result verification and iterative optimization. The overall process is clearly hierarchical, logically integrated, and interconnected with upstream and downstream processes. Specifically, it includes five core modules: a data layer, an intelligent parsing layer, an iterative optimization layer, an interpretable output layer, and an overall solution update layer. The correspondence between each module and the diagram, as well as their workflow, are as follows: First, the data layer corresponds to Figure 2The "Multi-Source Heterogeneous Data" module and the "Preprocessing" module on the left are responsible for aggregating multi-source heterogeneous data from annuity business scenarios. This includes structured data such as annuity account data, market transaction data, and macroeconomic indicators, as well as unstructured data such as business rule texts, industry dynamic information texts, and corporate financial report texts. Through differentiated and standardized preprocessing methods, the module completes data cleaning, normalization, and quantitative transformation, outputting high-quality standardized original basic annuity characteristic data, providing a stable and standardized data input foundation for subsequent intelligent analysis and computation.

[0023] Secondly Figure 2 Zhongda Model: Serving as the brain of the architecture, it comprises three customized models: an annuity operational characteristic data identification model, an annuity risk assessment model, and a configuration optimization model. Based on pre-training and fine-tuning technology, it adapts to annuity scenarios, achieving end-to-end intelligent decision-making for "annuity operational characteristic data identification, risk assessment, and configuration optimization." Specifically: Intelligent parsing layer corresponding Figure 2 The dual-branch parallel processing module of "Annuity Field Operational Feature Data Identification Model 1" and "Annuity Risk Assessment Model 2" adopts a dual-branch parallel processing architecture: based on standardized annuity original basic feature data, it extracts annuity field operational feature data and annuity risk feature data respectively, and then inputs the two types of feature data into the pre-trained annuity field operational feature data identification model and annuity risk assessment model respectively, to complete intelligent market situation identification and asset risk quantitative assessment in parallel, and output the corresponding annuity field operational feature data identification results and quantitative risk assessment results simultaneously, realizing full coverage mining of multi-dimensional deep features.

[0024] Iterative optimization layer corresponding Figure 2 The “Configuration Optimization Model 3 (based on Actor-Critic reinforcement learning algorithm)” module integrates three types of multi-dimensional data: original basic annuity feature data, annuity operational feature data identification results, and quantitative risk assessment results. These data are then input into the configuration optimization model. The model relies on the Actor-Critic reinforcement learning algorithm to conduct multiple rounds of constrained iterative optimization, adaptively generating annuity asset allocation ratio data that is in line with market dynamics and risk compliance.

[0025] Next, explain the corresponding output layer. Figure 2 The “First Attention Weight / Second Attention Weight” and the visualization output link at the top right are configured to simultaneously input the annuity asset allocation ratio data output by the optimized model into the large model. The large model outputs the first attention weight and the second attention weight, which are then aggregated to generate the total attention weight matrix, the aggregated weight of each dimension, and the weight ratio of each dimension. Finally, the visualization output of the annuity allocation decision explanation text is realized.

[0026] Finally, the overall solution update layer corresponds to Figure 2The monitoring link for "annuity asset allocation ratio data," the "annuity risk pressure verification and processing," and the dynamic fine-tuning link on the lower right are dynamically optimized through two mechanisms: First, the annuity risk pressure verification and processing performs risk compliance verification on the output results of the allocation optimization model, and unqualified results will trigger a re-iteration; Second, the business deviation monitoring and fine-tuning mechanism collects operational efficiency deviation data, annuity asset operation risk deviation data, and compliance execution data to trigger incremental fine-tuning of the allocation optimization model, regenerate annuity asset allocation ratio data adapted to industry and operational risk changes, and form a continuously optimized overall solution evolution capability.

[0027] This schematic diagram fully illustrates the overall technical architecture of this invention, which differs from traditional static, single-dimensional, and fixed processing modes. It clearly demonstrates the core technical principles of this invention: multi-source fusion, dual-path analysis, risk constraint, dynamic iteration, and overall scheme evolution. It intuitively explains the complete implementation path of this invention for achieving refined, intelligent, and dynamic processing of annuity asset data. The following section combines... Figure 2 A detailed introduction will be provided.

[0028] In step 101 above: In the initial stage of annuity asset data processing, the first step is to conduct comprehensive collection and standardized preprocessing of multi-source heterogeneous data from annuity business scenarios. The multi-source heterogeneous data collected in this embodiment consists of real business data derived from the entire annuity business chain, specifically covering two major types: structured data and unstructured data. Structured data mainly includes regular business data such as annuity account data, asset management transaction data, and macroeconomic indicator data, while unstructured data mainly includes irregular text data such as annuity business rule texts, industry dynamic information texts, and corporate operating financial statements. This achieves full coverage collection of annuity business-related data, ensuring the comprehensiveness of data sources and business adaptability.

[0029] After completing the collection of multi-source heterogeneous data, differentiated preprocessing operations were performed to address the structural differences between the two types of data, achieving standardization and organization of the messy raw data. For the collected structured data, data normalization, one-hot encoding conversion, and missing value interpolation were sequentially performed to unify the dimensions and units of measurement for various types of structured data, correct missing and anomaly issues in the raw data, and eliminate analytical interference caused by differences in the magnitude of different indicator data. For the collected unstructured text data, redundant information removal, text segmentation, and semantic embedding vectorization were sequentially performed to automatically filter invalid redundant fields, duplicate text, and meaningless interference information, transforming the unstructured text into high-dimensional semantic vector data that can be recognized and processed by large models.

[0030] Through the above differentiated and refined preprocessing operations, the unified standardization and feature purification of multi-source heterogeneous data are completed, eliminating problems such as messy original data formats, inconsistent dimensions, information redundancy, and missing data. Finally, standardized annuity original basic feature data with unified dimensions, effective features, and compliant quality are generated, providing standardized and reliable underlying data support for subsequent dual-branch feature extraction, large-scale model intelligent analysis, and dynamic iterative optimization calculation.

[0031] In one embodiment, the multi-source heterogeneous data includes structured data and unstructured data, as shown in Table 1 below. The structured data includes annuity account data, market transaction data, and macroeconomic indicators, while the unstructured data includes business rule text, industry dynamic information text, and corporate financial report text. Preprocessing of the multi-source heterogeneous data includes: performing normalization, one-hot encoding, and missing value interpolation on the structured data, and performing redundant information removal, word segmentation, and semantic embedding vectorization on the unstructured data.

[0032] In specific implementation, regarding Table 1 below, data acquisition adopts a distributed high-concurrency data acquisition engine. A single machine supports 1000~5000 concurrent connections, with a peak throughput of 3000~10000 TPS. It supports parallel acquisition of up to 64 heterogeneous data sources, and the concurrency capability can be linearly improved through cluster expansion. Engine fault tolerance mechanism: It adopts connection pool reuse, timeout control, circuit breaker protection, three exponential backoff retries, breakpoint resume, and master-slave second-level switching to ensure high availability of acquisition. Abnormal data processing logic: It adopts a hierarchical processing overall solution logic to detect, mark, filter, complete, or block anomalies such as missing, out-of-bounds, duplicate, and format errors. Abnormal data is fully retained and traceable. At the same time, it ensures data consistency through idempotency and deduplication to meet the high reliability requirements of financial-grade acquisition. Triggering conditions: Real-time acquisition on a trading day, batch acquisition at the end of the day; incremental acquisition is immediately triggered when events such as the release of coin business rules, significant market fluctuations, fund change data, and regulatory documents occur. Data flow: After acquisition, it is automatically pushed to the data processing layer.

[0033] Table 1: Multi-source heterogeneous data

[0034] In practice, the specific implementation methods for multi-source heterogeneous data acquisition and preprocessing are as follows: In one specific embodiment, this application completes the comprehensive collection and standardized preprocessing of multi-source heterogeneous data from annuity business scenarios in the initial stage of annuity asset data processing. This achieves the standardization, purification, and dimensional unification of the original business data, providing high-quality and standardized basic data support for subsequent large-scale model feature extraction, intelligent analysis, and iterative optimization. Specifically, the collected multi-source heterogeneous data includes both structured and unstructured data. These two types of data have complementary sources and different feature dimensions, together forming a complete data foundation for annuity asset data analysis. Specifically, the structured data includes annuity account data, market transaction data, and macroeconomic indicators. This type of data is characterized by high standardization, clear dimensions, and direct quantification, and can intuitively reflect the basic status of annuity accounts, asset management transaction dynamics, and macroeconomic environment changes. The unstructured data includes business rule text, industry dynamic information text, and corporate financial report text. This type of data is natural language text data without a fixed format, which can fully cover annuity business regulations, industry dynamic information, and corporate fundamentals, compensating for the insufficient coverage of scenario and industry information dimensions by structured data.

[0035] To address the structural differences and data attributes of the two types of heterogeneous data, this application adopts a differentiated preprocessing strategy, performing corresponding standardized processing operations to achieve unified organization and feature purification of the messy original data. Specifically, the collected structured data undergoes normalization, one-hot encoding, and missing value interpolation. Normalization unifies the dimensions and numerical ranges of various types of structured data, eliminating analytical interference caused by differences in magnitude between different indicators; one-hot encoding digitizes the discrete features of the structured data to adapt to the model's input requirements; and missing value interpolation intelligently completes blank and missing fields in the original data, avoiding model analysis bias caused by data deficiencies and ensuring the integrity and standardization of the structured data.

[0036] Correspondingly, the collected unstructured data undergoes redundant information removal, word segmentation, and semantic embedding vectorization. Specifically, the following steps are taken: First, redundant information is removed from the original text data, automatically filtering out invalid characters, duplicate content, irrelevant public opinion noise, and meaningless interference information, while retaining core business and industry-specific information. Then, word segmentation is performed on the purified standard text content, dividing the continuous natural language text into independent and effective semantic vocabulary units. Finally, through semantic embedding vectorization, the discrete text semantic information is transformed into high-dimensional semantic vectors with unified dimensions that can be recognized and processed by large models, thereby achieving the structured and digital transformation of unstructured text data.

[0037] Through the aforementioned differentiated and refined preprocessing operations, this application completes the cleaning, normalization, and feature quantification of multi-source heterogeneous data, effectively solving the technical problems of messy original annuity business data formats, inconsistent dimensions, redundant information, missing data, and inability to model and perform calculations on text. It obtains standardized original basic annuity feature data with unified dimensions, effective features, and stable quality, providing reliable and complete data support for subsequent dual-branch feature extraction, intelligent model judgment, and dynamic iterative optimization.

[0038] In practical implementation, data preprocessing serves as a "bridge between raw data and the large model," its core function being to transform heterogeneous data into a standardized format that the large model can parse, eliminating potential data quality issues. For the collected structured data, this layer uses a normalization algorithm to map values ​​to a unified range, employs one-hot encoding to process categorical information (such as bond ratings), and uses interpolation to fill in missing values, ensuring quantitative consistency. For unstructured data, redundant information is first removed, and then semantic embedding technology is used to convert text into vector form, retaining core semantic features such as annuity business rule keywords and public opinion trends. Simultaneously, this layer integrates a data anonymization module to encrypt sensitive account information, meeting both the input requirements of the large model and complying with financial data security standards, providing the large model with "high-quality and high-security" input data. Triggering condition: Triggered immediately upon arrival of data from the data acquisition layer, with streaming processing and real-time input into the model. Data flow: After processing, data is uniformly input into the core layer of the large model. Data anonymization: A combination of national cryptographic standard SM4 encryption, order-preserving mapping anonymization, and differential privacy perturbation is used to securely process sensitive data such as account identifiers, transaction amounts, and customer information. While hiding sensitive information, the data statistical features and semantic information are preserved, which not only meets the input requirements of large models, but also complies with financial data security and regulatory compliance standards, providing "high-quality and high-security" input data for large models.

[0039] In step 102 above: After acquiring the original basic characteristic data of standardized annuities, a dual-branch parallel feature extraction and intelligent model parsing process is executed. Two dedicated pre-trained models are used to complete business situation identification and risk quantification assessment, respectively, achieving comprehensive and parallel mining and analysis of the deep correlation features of the original annuity data. This step adopts a dual-branch independent computation and synchronous processing technical architecture, breaking the limitations of the traditional single-path serial data processing mode. It can fully reuse all effective information from the original basic characteristic data of standardized annuities, simultaneously completing the analysis and judgment output of data features from different dimensions.

[0040] The first approach involves identifying operational characteristics in the pension sector. This process involves precisely filtering and extracting operational characteristics relevant to market analysis from pre-processed raw pension data. These characteristics encompass multi-dimensional information, including market transaction volatility, macroeconomic correlations, and industry operational dynamics. The extracted operational characteristics are then input into a trained pension sector operational characteristics identification model. Leveraging the model's built-in feature correlation reasoning capabilities, the model quantitatively analyzes the current market dynamics related to pension business, outputting standardized operational characteristics identification results. This enables accurate perception of the external market environment and business operational trends.

[0041] The second approach involves a quantitative risk assessment process for annuities. This process simultaneously extracts annuity risk characteristic data tailored to the risk assessment dimensions from the original basic annuity data. This data covers core risk dimensions such as annuity account operational risk characteristics, asset volatility risk characteristics, and potential business operation risk characteristics. The extracted annuity risk characteristic data is then input into the annuity risk assessment model. Through correlation analysis, weight matching, and quantitative extrapolation of multi-dimensional risk characteristics, the model accurately identifies various potential risk factors in the annuity business operation process, quantifies the risk level and volatility, and ultimately outputs standardized quantitative annuity risk assessment results.

[0042] In practical implementation, regarding the data identification model for operational characteristics in the annuity field: This model corresponds to the first branch of the dual-branch intelligent analysis architecture: the annuity sector operational feature identification process. The overall execution logic is as follows: From the pre-processed raw annuity feature data, it accurately filters and extracts annuity sector operational feature data that matches the market analysis dimensions. This type of feature data covers multi-dimensional related information such as market transaction fluctuation characteristics, macroeconomic correlation characteristics, and industry operational dynamic characteristics. Subsequently, the extracted annuity sector operational feature data is input into the trained annuity sector operational feature data identification model. Relying on the model's built-in feature association reasoning capabilities, it quantitatively analyzes the current market operation status related to annuity business, outputting standardized annuity sector operational feature data identification results, thus achieving accurate perception of the external market environment and business operation status.

[0043] The annuity sector operational characteristic data identification model is built upon BERT and LLM pre-trained large models and the Transformer architecture, specifically developed for annuity asset data feature mining and annuity sector operational analysis scenarios. The model relies on existing historical annuity sector operational data, macroeconomic data, and annuity business rule texts to construct a dedicated training set (containing over 1 million samples). It employs a mini-batch gradient descent method for domain-specific fine-tuning, enabling deep understanding and analysis of complex financial texts and time-series market data. Through training with massive amounts of financial sector-specific data, it accurately mines financial market operation patterns, multi-dimensional data correlations, and time-series evolution characteristics. It completes annuity sector operational pattern analysis, data correlation analysis, and quantitative judgment of market operation status, assisting in the identification of market operation characteristics, industry prosperity characteristics, and macroeconomic correlation characteristics. This provides standardized feature basis for subsequent risk assessment and allocation optimization, without involving subjective predictions, only completing the technical data for data feature mining and state quantitative analysis.

[0044] In practical implementation, the base model for identifying operational characteristics data in the annuity field can adopt BERT-base-chinese, with a 12-layer Transformer encoder, 12-head self-attention, 768-dimensional hidden layers, and a 512-dimensional context window. The model training uses over 1 million labeled samples, covering nearly 10 years (2016-2026); the data types and proportions are as follows: macroeconomic indicators 30%, market transaction data 35%, annuity business rules text 25%, and industry dynamics and analyst reports 10%; the sample selection criteria are: compliant and authoritative sources (third-party public data interfaces, etc.), highly relevant to annuity asset operations, no missing or abnormal data, and labeled with trends for the next 1-6 months.

[0045] The annuity field's operational feature data identification model can be fine-tuned using mini-batch gradient descent. The key training parameters are: batch size = 32, learning rate = 2e-5, training epochs = 10, the optimizer is AdamW, and the loss function combines cross-entropy loss and MSE time-series regression loss, with dropout = 0.1 and weight decay added.

[0046] In specific implementation, the pre-establishment process of the annuity field operation characteristic data identification model is as follows: 1) Collect financial annuity business rules, macro indicators, and market text corpora in the annuity field; 2) Load BERT pre-trained weights; 3) Conduct secondary pre-training using unlabeled corpora in the field; 4) Construct a trend-labeled sample set; 5) Perform fine-tuning training according to the above parameters and select the best in the validation set; 6) Perform knowledge distillation and model compression to meet real-time requirements; 7) Deploy and access the collaborative scheduling module to provide online inference services.

[0047] In practice, the detailed working process of the annuity field operational feature data identification model is as follows: First, the input macroeconomic data and market data are normalized, and the annuity business rule text is segmented and embedded with encoding, and time-series position encoding is uniformly added; then, through 12 layers of Transformer self-attention encoding, the keywords, indicator correlations and time-series trend features of the annuity business rules are extracted; the global semantic vector is obtained through pooling, and then the classification result is output through a fully connected layer, providing trend basis for risk assessment and configuration optimization.

[0048] In practice, the input to the annuity field operation characteristic data identification model can be annuity field operation characteristic data, and the output can be the annuity field operation characteristic data identification result.

[0049] In practical implementation, regarding the annuity risk assessment model: The annuity risk assessment model is the core of this invention's security and control system. It is constructed using a multi-task learning framework, embedding annuity-specific risk factors (such as annuity business rule compliance, credit default, and interest rate volatility). Each factor corresponds to an independent sub-model branch, quantifying and calculating asset risk values ​​and high-risk points. Finally, a weighted fusion is used to output a comprehensive risk value. The specific implementation process is as follows: Five key annuity-specific risk factors, namely annuity risk characteristic data.

[0050] 1) Sub-model for compliance risk of annuity business rules: Determine whether the allocation ratio, asset operation scope, and product access comply with regulatory rules (quantitative calculation formula: (number of non-compliant items × weight + excess ratio × coefficient) × 100; if prohibited asset operation is involved, 100 points will be directly deducted). 2) Credit Default Risk Sub-model: Assess the probability of default (PD) and loss given default (LGD) of bonds, fixed income and non-standard assets (quantitative calculation formula: PD×LGD×100); 3) Interest Rate Volatility Risk Sub-model: Calculates the impact of interest rate changes on bond market value, portfolio duration, and net asset value volatility (quantitative calculation formula: |duration × interest rate change in BP| × conversion factor × 100). 4) Liquidity Risk Sub-model: Assess the liquidity of holdings, the cost of portfolio rebalancing, and the daily redeemable size (quantitative calculation formula: (1 / average daily turnover rate) × holding ratio × 100); 5) Concentration risk sub-model: Monitor whether the holdings of industries, varieties, issuers, and single accounts exceed the limits (quantitative calculation formula: holding ratio of a single industry / variety / issuer × 100).

[0051] Weighting and weighted fusion

[0052] 1) Benchmark weights: Annuity business compliance 30%, credit default 25%, interest rate volatility 20%, liquidity 15%, concentration 10%.

[0053] 2) Dynamic adjustment: The weight of annuity business rules is increased during the annuity business rule cycle, the weight of liquidity is increased during periods of sharp market fluctuations, and the weight of interest rate risk is increased during interest rate cycles.

[0054] 3) Overall Risk Score: Σ (sub-factor score × dynamic weight), outputting the overall risk value, risk level, potential risk range, and upper limit of constraints.

[0055] Model pre-establishment process

[0056] 1) Identify five major risk factors based on annuity regulatory rules; 2) Construct a risk feature database; 3) Collect over 800,000 labeled samples from the past 10 years (2016-2026); 4) Design a multi-branch sub-model structure; 5) Train using the AdamW optimizer, batch size=32, learning rate 1e-4, epoch=8; 6) Learn the optimal weighting coefficients; 7) Calibrate the risk level threshold; 8) Deploy and access collaborative scheduling.

[0057] In practice, the input to the annuity risk assessment model can be annuity risk characteristic data, and the output can be the annuity quantitative risk assessment result.

[0058] It should be noted that both types of intelligent analysis models used in this application are pre-trained models specifically for annuity-specific business scenarios, possessing strong scenario adaptability and analytical accuracy. Specifically, the training samples for the annuity operational characteristic data identification model consist of massive historical annuity operational characteristic data and corresponding standard identification results. Through multiple rounds of sample iterative training, parameter optimization, and accuracy verification, it fully learns the correlation patterns between annuity market characteristics and trend changes, possessing accurate quantitative identification capabilities for market trends. The annuity risk assessment model, on the other hand, relies on massive historical annuity risk characteristic data and corresponding standardized quantitative risk assessment result samples for pre-training. It deeply fits the risk generation logic and volatility characteristics of annuity business, accurately adapting to the specific risk assessment needs of annuity business and effectively avoiding the problems of poor adaptability and low assessment accuracy of general models.

[0059] In step 103 above: After completing the standardized preprocessing of multi-source data and the parallel intelligent analysis and judgment of dual branches to obtain complete multi-dimensional data and analysis results of annuity business, further multi-dimensional data fusion and intelligent iterative optimization processing are carried out to achieve in-depth processing and adaptive judgment upgrades of annuity asset data. This step relies on a reinforcement learning iterative optimization mechanism to break the traditional static, single, and fixed operation mode of annuity data processing, and integrates multi-dimensional core data for linkage analysis and dynamic optimization calculations, effectively improving the intelligence level and accuracy of annuity asset data processing.

[0060] Specifically, the standardized basic annuity characteristic data generated through preprocessing, the annuity operational characteristic data identification results output by the annuity operational characteristic data identification model, and the annuity quantitative risk assessment results output by the annuity risk assessment model—three types of core data with different dimensions and attributes—are fully integrated and uniformly input into a pre-set optimization model. The original basic annuity characteristic data retains all the underlying effective characteristics of annuity business; the annuity operational characteristic data identification results reflect the current operational dynamics and environmental characteristics of annuity-related businesses; and the quantitative risk assessment results accurately characterize the risk level and volatility characteristics of annuity assets. These three types of data complement and corroborate each other, providing comprehensive and multi-dimensional data support for subsequent iterative optimization calculations, avoiding the analytical bias caused by a single data dimension.

[0061] In this embodiment, the optimization model incorporates reinforcement learning iterative optimization logic. With the optimization goal of adapting to the dynamic operational characteristics of annuity business and meeting the needs of annuity asset data processing and analysis, it performs multiple rounds of iterative calculations and adaptive logic tuning on the input multi-dimensional fused data. During the iteration process, the model continuously explores the deep correlations between underlying basic data, market situation data, and risk data, dynamically corrects data analysis weights and logic, and continuously optimizes the data processing results, abandoning the traditional static calculation mode with fixed parameters and formulas. Through multiple rounds of iterative optimization and continuous convergence, it adaptively completes the intelligent and in-depth processing of annuity asset data, ultimately outputting accurate, stable, and adaptable annuity asset data analysis results suitable for complex business scenarios. This achieves end-to-end intelligent processing of annuity asset data, from basic processing and intelligent parsing to dynamic optimization and analysis, significantly improving the refinement level and adaptability to complex scenarios in annuity asset data processing.

[0062] In one embodiment, in step 103 above, the configuration optimization model uses the original basic characteristic data of annuities and the identification results of operational characteristic data in the annuity field as input data to complete the deduction calculation, and uses the quantitative risk assessment results of annuities as optimization constraints to complete multiple rounds of iterative optimization based on the Actor-Critic reinforcement learning algorithm.

[0063] In practice, the iterative optimization method based on Actor-Critic reinforcement learning combined with dynamic optimization of a large model is as follows: In one specific embodiment, the configuration optimization model adopted in this application relies on the Actor-Critic reinforcement learning algorithm combined with the contextual reasoning capability of a large model to achieve intelligent, constrained, and dynamic multi-round iterative optimization deduction, comprehensively revolutionizing the static processing drawbacks of traditional annuity asset allocation models that rely on manually setting parameters, manually defining data correlations, and manually defining constraints. Traditional technical solutions mostly adopt a linear operation mode with fixed formulas and static parameters, resulting in fixed parameters, poor adaptability, and an inability to adjust asset allocation in real time with market and risk changes. This invention, through the dual-drive mechanism of Actor-Critic reinforcement learning autonomous iterative optimization and large-model contextual intelligent reasoning, completes the dynamic optimization and real-time rebalancing of annuity asset data. The overall model is based on multi-dimensional real business data and uses quantified risk results as rigid constraint boundaries, ensuring the accuracy, real-time performance, and compliance of the data processing results.

[0064] The specific implementation process is as follows: In the model iteration calculation stage, this embodiment strictly limits the input data source and constraints of the model. The standardized original basic feature data of annuities and the identification results of the annuity field operation feature data obtained by parallel analysis of dual models are used as the core input data for configuring and optimizing the model. This provides complete underlying business features and market situation features to support the model inference calculation, ensuring that the model calculation has comprehensive, real and real-time data basis. At the same time, the annuity quantitative risk assessment results output synchronously are set as the core constraints for model iteration optimization. The direction of model iteration is controlled by quantitative risk indicators throughout the process to avoid problems such as risk deviation and scenario adaptation failure in the optimization calculation process.

[0065] This invention employs the classic Actor-Critic dual-body reinforcement learning algorithm to complete multiple rounds of iterative optimization. The overall system consists of two main functional modules: an Actor policy network and a Critic evaluation network, working collaboratively. Combined with the powerful contextual understanding and dynamic reasoning capabilities of a large model, it achieves autonomous optimization of parameters and allocation ratios without manual intervention. Specifically, the Actor policy network is responsible for identifying the results of input annuity basic feature data and annuity operational feature data, continuously outputting annuity asset data optimization strategies adapted to the current business scenario and market dynamics, and generating initial asset allocation projection results. The Critic evaluation network, using the annuity quantitative risk assessment results as the core constraint benchmark, combines historical business data and risk threshold standards to perform real-time evaluation, scoring, and deviation judgment on the projection results output by the Actor policy network, accurately identifying risk deviations and scenario adaptation deviations in the current optimization strategy.

[0066] During multiple rounds of iterative optimization, the model uses the feedback error value from the Critic evaluation network to correct the internal weight parameters and strategy generation logic of the Actor strategy network, continuously optimizing the asset allocation deduction strategy and narrowing the deviation between the deduction results and business compliance and risk control standards. Simultaneously, leveraging the contextual reasoning capabilities of the large-scale model, it captures real-time market dynamics, business rule updates, and risk situation fluctuations, assisting the reinforcement learning model in adaptive parameter fine-tuning and achieving real-time rebalancing of asset allocation. Through iterative cycles of dynamic perception by the large-scale model, Actor strategy generation, Critic risk assessment, and reverse parameter updates, multiple rounds of adaptive optimization are achieved, gradually converging to obtain the optimal annuity asset allocation ratio data that aligns with market conditions, matches business characteristics, and meets risk constraints.

[0067] In practical implementation, regarding the configuration optimization model: This reinforcement learning and large-scale model-based constrained dynamic iterative computation mode enables two-way linkage between forward data extrapolation, reverse risk constraints, and dynamic scenario adaptation. It effectively solves the technical pain points of traditional manual parameter tuning mode, such as fixed parameters, simple optimization logic, lack of risk constraints, and inability to dynamically adapt to changes in market and business risks. It significantly improves the robustness, accuracy, and real-time scenario adaptability of annuity asset data processing results.

[0068] In specific implementation, this invention integrates reinforcement learning (RL) with large model inference. It iteratively optimizes the asset allocation ratio through the Actor-Critic algorithm (At=rt+γV(st+1)-V(st)). At the same time, it introduces the contextual understanding capability of the large model to adapt to dynamic constraints (such as regulatory ratio limits and customer risk preferences). It adopts an offline training + online fine-tuning mode. In the offline stage, the basic strategy is trained based on historical data. In the online stage, it is adjusted and iterated in real time through supervision feedback data (such as updating the training set every quarter) to ensure that the strategy adapts to market changes.

[0069] In practice, the triggering conditions are: data processing is completed, market fluctuations exceed the threshold, risk indicators exceed limits, the scheduled rebalancing cycle is reached, and the monitoring feedback layer triggers a model update.

[0070] In practical implementation, the sub-models (annuity field operational characteristic data identification model, annuity risk assessment model, and configuration optimization model) and their scheduling mechanism can utilize a unified scheduling module (parallel parsing unit and processing unit) to employ a unified scheduling method of "parallel triggering, serial dependency, and result fusion" for the three sub-models. 1) The processed data is simultaneously distributed to the annuity field's characteristic data identification model and annuity risk assessment model, and the two are calculated in parallel and independently; 2) After both outputs are complete, merge the original basic characteristic data of annuities, the identification results of operational characteristic data of annuities, and the quantitative risk assessment results of annuities into the configuration optimization model; 3) Configure and optimize the large model to be executed serially to generate the final asset allocation ratio.

[0071] Sub-model functions and linkage relationships

[0072] 1) The annuity field operation feature data identification model is based on BERT, LLM pre-trained large model and Transformer architecture. It uses existing historical annuity field operation data, macroeconomic data and annuity business rule text to build a dedicated training set (containing more than 1 million samples). It is fine-tuned by mini-batch gradient descent method to understand and process complex financial text information and can accurately analyze the results of the financial annuity field operation feature data identification.

[0073] 2) The annuity risk assessment model (core of security and control) is based on a multi-task learning framework, embedding annuity-specific risk factors (annuity business rule compliance, credit default, interest rate fluctuations, liquidity risks, etc.). Each factor corresponds to an independent sub-model branch, quantitatively calculating asset risk values ​​and high-risk points, and finally weighted and fused to output a comprehensive risk value. Output: Comprehensive risk level, potential risk range, compliance risk warning, i.e., the annuity quantitative risk assessment result. Interrelationship: The risk factor values ​​are dynamically affected by the identification results of annuity operational characteristic data, directly changing the annuity quantitative risk assessment result.

[0074] 3) Configuration optimization of the large model (core of strategy generation): It integrates reinforcement learning (RL) and large model inference, and iteratively optimizes the asset allocation ratio through the Actor-Critic algorithm. At the same time, it introduces the contextual understanding ability of the large model to adapt to dynamic constraints (such as regulatory ratio limits and customer risk preferences). It adopts offline training and online fine-tuning modes. In the offline stage, the basic strategy is trained based on historical data. In the online stage, it is adjusted and iterated in real time through supervised feedback data (such as updating the training set every quarter) to ensure that the strategy adapts to market changes.

[0075] Linkage: It must be based on the identification of operational characteristics data in the annuity field and risk assessment as a hard constraint. The configuration ratio cannot be generated independently. The output configuration ratio is sent back to the risk assessment model for stress verification. If the threshold is not met, it is iterated again.

[0076] Further preferred embodiments of the present invention are described below.

[0077] Figure 3 This is a schematic diagram of the process for visualizing the output of decision explanation text in one embodiment of the present invention. In one embodiment, such as... Figure 3 As shown, the above-mentioned method for processing annuity asset data based on a large model may also include the following steps: Step 201: Extract the first classification label, which serves as the global feature aggregation identifier, from the preset layer network of the annuity field operation feature data identification model, relative to the original basic feature data of annuities, the annuity field operation feature data identification results, and the annuity quantitative risk assessment results; Step 202: Extract the second classification label, which serves as the global feature aggregation identifier, from the pre-defined layer network of the annuity risk assessment model, relative to the identification results of the original basic annuity feature data, the annuity domain operational feature data, and the annuity quantitative risk assessment results, as the second attention weight; Step 203: Combine the first attention weight and the second attention weight to generate a total attention weight matrix; Step 204: Aggregate the total attention weight matrix according to four dimensions: business rule keywords, macro indicators, domain operation data, and annuity-specific risk factors, to obtain the aggregated weights for each dimension; Step 205: Based on the aggregated weights of each dimension, the aggregated weight of a single dimension is calculated as a ratio to the total weight of all dimensions using a preset normalization calculation method, and then converted into a percentage to obtain the weight percentage of each dimension. Step 206: Complete the visualization rendering based on the weight ratio of each dimension. Each set of annuity asset allocation ratio data corresponds to a unique dimension weight distribution feature. Generate matching annuity allocation decision explanation text based on the weight distribution feature to realize the visualization output of the annuity allocation decision explanation text.

[0078] In practice, the visualization and decision interpretation steps can be implemented as follows: In terms of visualizing the explanation text of pension allocation decisions, this invention, in one specific embodiment, further adds a model feature attribution analysis and visualization explanation process based on the pension asset data processing results after iterative optimization. This achieves traceability, quantification, and displayability of the intelligent data processing process. The specific implementation steps are as follows: First, a dual-model attention weight extraction operation is performed, and deep network weight analysis is conducted for the pension field operational feature data identification model and the pension risk assessment model, respectively. Specifically, the first classification label used for global feature aggregation in the preset network layers of the pension field operational feature data identification model is selected, and the correlation contribution of the first classification label relative to the three core data types—the original basic feature data of pensions, the pension field operational feature data identification results, and the pension quantitative risk assessment results—is calculated to obtain the corresponding first attention weight. Simultaneously, the second classification label that plays a role in global feature aggregation in the preset network layers of the pension risk assessment model is selected, and the correlation influence weight of the second classification label relative to the above three core data types is measured to obtain the second attention weight. Through dual-model hierarchical weight extraction, the deep feature contribution patterns of the two models in the data analysis and feature judgment process are accurately captured.

[0079] After obtaining the attention weights corresponding to the two models, matrix alignment and fusion operations are performed on the first and second attention weights to unify the weight dimensions and units, eliminate weight bias caused by differences in network parameters of different models, and generate a total attention weight matrix covering the entire model and all data dimensions. This process fully preserves all feature-related weight information from the parallel analysis of the two models and the multi-dimensional data fusion analysis. Subsequently, multi-dimensional weight aggregation processing is carried out. Based on the core influencing factors of annuity asset management business, this embodiment defines four analytical dimensions: business rule keywords, macro indicators, domain operation data, and annuity-specific risk factors. The scattered single-point weight data in the total attention weight matrix are categorized, superimposed, and aggregated according to the four business dimensions to obtain the corresponding dimension aggregation weights. This achieves the business-oriented and structured integration of scattered weight data, accurately collecting the overall contribution value of different business dimensions to the data processing results.

[0080] Based on this, the weight proportion is quantified through a preset normalization calculation method. First, the aggregated weights of the four dimensions are summarized to obtain the global total weight. Then, the aggregated weight of each individual dimension is compared with the global total weight and converted into a percentage of the dimension weight proportion. Normalization eliminates differences in data magnitude, intuitively presenting the relative contribution of each business dimension to the annuity asset data processing results. Finally, visualization rendering and intelligent explanatory text generation are carried out. Based on the weight proportion data corresponding to each dimension, a unique dimension weight distribution feature map is constructed to complete the data visualization rendering. Each set of annuity asset data processing results corresponds to a unique dimension weight distribution feature. At the same time, based on preset text generation rules and weight distribution features, corresponding annuity data analysis and explanation text is automatically generated, clearly marking the core influencing dimensions, weight proportion distribution, and data analysis basis. Ultimately, the visualization and textual output of the annuity asset data processing process, feature attribution logic, and analysis results are realized, completing the entire process of interpretable data processing.

[0081] In practical implementation, this invention extracts, merges, aggregates, and quantifies the attention weights of the two large-scale model network layers. Combined with multi-dimensional weight proportions, it achieves visualized rendering and automatic generation of decision explanation text. This effectively overcomes the technical shortcomings of traditional large-scale model annuity asset data processing solutions, such as black-box computation, unquantifiable feature contributions, and poor interpretability of analysis results. Traditional intelligent annuity data processing methods can only output the final judgment result, failing to trace the impact of different business dimensions on the analysis result. It is difficult to intuitively reflect the differentiated contribution characteristics of the four core dimensions: macro indicators, domain operational data, business rules, and risk factors, resulting in weak persuasiveness, opaque processes, and insufficient traceability of data processing results. This solution extracts the global features of the two dedicated large-scale models, aggregates the attention weights, and merges them to generate a unified total attention weight matrix, achieving accurate quantification and unified representation of the deep feature contributions of the model. Simultaneously, it aggregates weights and calculates normalized ratios according to the core dimensions of annuity business, accurately quantifying the true impact proportion of various data dimensions on the final data processing result, overcoming the limitations of traditional technologies that cannot quantify dimension contributions and have ambiguous feature attributions. Building upon this foundation, the system leverages dimensional weight distribution characteristics to automatically generate visual rendering and customized decision explanation texts. This transforms the previously abstract and black-boxed intelligent data processing process into a visual, traceable, and interpretable quantitative basis, clearly presenting the generation logic and core influencing factors of annuity asset data processing results. This significantly enhances the transparency, interpretability, and business implementation credibility of the intelligent data processing process for annuity assets. It provides intuitive and accurate quantitative support and visual evidence for annuity asset data analysis, business review and verification, and traceability of intelligent analysis results, further improving the overall solution's practicality and adaptability in complex annuity business scenarios.

[0082] In practice, the decision interpretation document module utilizes large-scale model attention visualization technology to display key influencing factors in decision-making (such as the weighting of keywords in annuity business rules). The output is directly pushed to the asset management manager's terminal and the regulatory platform, supporting business execution while meeting regulatory transparency requirements.

[0083] Triggering conditions: The core layer of the large model outputs decision results, risk signals are generated, and the configuration ratio changes.

[0084] Data flow: Output usable business results and send them to the supervision and monitoring layer simultaneously.

[0085] Configuration scheme generation module implementation mechanism: (1) Basic weights: taken from the optimal asset allocation output by the large-scale configuration optimization model Actor-Critic; (2) Adjustment item: Based on the trend confidence of the annuity field operation characteristic data identification model, the adjustment is made dynamically within ±5%.

[0086] Decision Explanation Document Module Implementation Mechanism: 1) Attention weight extraction: Extract the attention weight matrix of the last 2-4 layers of classification labels to each input unit from the annuity domain feature data recognition model.

[0087] 2) Dimensional aggregation: Weighted aggregation is performed based on dimensions such as annuity business rule keywords, macroeconomic indicators, field operation data, and risk factors.

[0088] 3) Weight Quantification Calculation: The weight ratio of keywords in annuity business rules is calculated using a normalization method.

[0089] 4) Keyword weight percentage = (Average attention weight of a single keyword / Total weight of all keywords) × 100%.

[0090] 5) Visual output: Presented in the form of weighted bar charts, keyword clouds, and impact factor rankings, and automatically generated decision explanation text. For example, in the annuity business rules, the keyword "stable growth" accounts for 30% of the weight, which is the core influencing factor in this configuration.

[0091] In one embodiment, the above-mentioned large-model-based annuity asset data processing method may further include: The generated annuity asset allocation ratio data is sent back to the annuity risk assessment model for annuity risk pressure verification, and the overall risk indicators of the current annuity asset allocation ratio data are checked to see if they meet the preset safety threshold. If the risk indicators do not meet the preset safety threshold, the current annuity asset allocation ratio data will be fed back to the allocation optimization model, and multiple rounds of iterative optimization will be performed again until the asset operation risk indicators meet the preset safety threshold, and compliant annuity asset allocation ratio data will be output.

[0092] In practice, the overall scheme for risk feedback verification and iterative optimization can be implemented as follows: In one specific embodiment, after completing multi-dimensional data fusion and iterative optimization and generating annuity asset data processing results, this application adds a risk overall solution verification and secondary iterative optimization mechanism. Through result feedback verification, it achieves an intelligent overall data processing solution with controllable risk and compliant results, effectively ensuring that the final output annuity asset assessment results are risk-controllable and conform to business security standards. The specific implementation process is as follows: This application backfeeds the annuity asset allocation ratio data adaptively generated by the configuration optimization model to the pre-trained annuity risk assessment model. Relying on the annuity risk assessment model's dedicated risk assessment capabilities, it conducts comprehensive annuity risk pressure verification processing for the current asset allocation structure. During the verification process, the model combines the original basic annuity characteristic data, real-time business situation characteristics, and historical risk patterns to comprehensively calculate core risk parameters such as the overall risk volatility index, pressure risk coefficient, and extreme scenario volatility risk corresponding to the current allocation structure. It comprehensively checks whether the comprehensive risk index corresponding to the current annuity asset allocation ratio data meets the business's preset safety threshold standard, achieving post-optimization risk compliance verification of the optimization results.

[0093] After verification, tiered judgment and overall solution iterative processing are performed. If the annuity risk assessment model verifies that the current overall risk indicators meet the preset safety threshold requirements, the current annuity asset allocation ratio data is deemed compliant with risk and meets business adaptability standards, and the compliant annuity asset data processing results can be directly output. If the various risk indicators obtained from verification do not meet the preset safety thresholds, and there are situations where risk pressure exceeds the standard, structural risk is too high, or the requirements for sound annuity operation are not met, the current non-compliant annuity asset allocation ratio data will be immediately fed back to the allocation optimization model, triggering the model's second iteration optimization mechanism. Based on the risk exceeding feedback information, the allocation optimization model will adaptively adjust the multi-dimensional data weight ratio and iterative optimization strategy, and re-rely on reinforcement learning algorithms to conduct multiple rounds of iterative optimization calculations, dynamically correcting and optimizing the asset allocation structure, and then feeding it back to the annuity risk assessment model for a new round of risk pressure verification.

[0094] This embodiment employs a cyclical overall solution mechanism of "optimized generation, risk feedback verification, and non-compliance re-iteration" to continuously iterate and correct data processing results until the overall risk indicators corresponding to the final output annuity asset allocation ratio data fully meet the preset safety threshold. At this point, the iteration process terminates, and the final compliant and robust annuity asset data processing result is output. This overall solution verification and iteration mechanism solves the problem that single-iteration optimization results are prone to insufficient risk adaptability and inability to match the requirements of low-risk and stable annuity operations. It ensures that the risk and compliance of annuity asset data processing results are controllable, further improving the security, stability, and reliability of intelligent annuity asset data processing results for business implementation.

[0095] In one embodiment, the above-mentioned large-model-based annuity asset data processing method may further include: Monitor the actual implementation results of annuity asset allocation ratio data, and collect data on operational efficiency deviations, annuity asset operation risk deviations, and compliance execution data. When various deviation data exceed the corresponding preset thresholds, the incremental fine-tuning of the configuration optimization model is triggered, the parameters of the configuration optimization model are updated, and the annuity asset allocation ratio data adapted to changes in industry and operational risks is regenerated.

[0096] In practice, the specific implementation methods for monitoring execution results and incremental model fine-tuning can be as follows: In one specific embodiment, after completing the compliant output of annuity asset data, this application adds a full-cycle execution monitoring and model incremental fine-tuning mechanism to achieve dynamic iterative updates and long-term adaptive adaptation of the data processing model, solving the technical problem of traditional static models being rigid and unable to adapt to the dynamic evolution of business. The specific implementation process is as follows: This application continuously and dynamically monitors the actual implementation effect of the output annuity asset allocation ratio data, tracks the actual operating status of annuity assets in real time, and systematically collects various types of deviation data and compliance data generated during the execution process, specifically including operational efficiency deviation data, annuity asset operation risk deviation data, and compliance execution data. Among them, operational efficiency deviation data is the difference between the expected return judged by the model and the actual return; operational risk deviation data is the deviation between the preset risk quantification indicators and the actual asset operation risk fluctuations; and compliance execution data is the compliance record data of whether the asset execution process conforms to the annuity business regulatory rules and business operation norms. Through multi-dimensional data collection, the true state of the actual operation of the assets is fully restored, providing a real and effective sample basis for model fine-tuning.

[0097] After completing routine data collection and statistics, the system compares various deviation data with corresponding preset deviation thresholds to accurately identify the degree of deviation between the actual execution effect and the theoretical output effect of the model. When any one or more of the collected profit deviation, risk deviation, or compliance execution deviation data exceed the corresponding preset threshold, it is determined that the adaptability of the current model parameters has decreased and the original data processing logic can no longer adapt to the current business and market dynamics, immediately triggering the incremental fine-tuning mechanism of the configuration optimization model. This embodiment uses incremental fine-tuning to perform local iterative updates to the model parameters. Unlike retraining the entire model, it only relies on the latest real execution deviation samples, compliance operation samples, and dynamic risk samples to finely adjust and update the internal weight parameters, iterative optimization strategies, and feature matching logic of the configuration optimization model. While retaining the model's original mature basic parameters and business feature learning capabilities, it corrects the model's lag deviation.

[0098] After the incremental update of model parameters is completed, the updated configuration optimization model, combined with the latest business situation, risk characteristics, and actual execution patterns, conducts multiple rounds of iterative optimization calculations to adaptively generate new annuity asset allocation ratio data that is adapted to the current dynamic market environment, risk volatility characteristics, and actual business execution patterns. This completes the dynamic update of data processing logic and output results. This mechanism, through a dynamic optimization mode of real-time monitoring, deviation triggering, incremental fine-tuning, and iterative updates, enables the continuous evolution of model capabilities. It effectively solves the problems of static solidification, long-term adaptation lag, and disconnect between output results and actual business in traditional annuity data processing models. This ensures that the annuity asset data processing logic always aligns with the actual business operation patterns, significantly improving the accuracy, timeliness, and dynamic adaptability of data analysis in long-term complex business scenarios.

[0099] like Figure 2 The "monitoring" layer serves as the "overall solution optimization guarantee" for the system. The supervisory feedback layer and the large model fine-tuning module jointly construct a "decision-making, execution, evaluation, and optimization" loop mechanism to ensure the large model's performance continuously adapts to market changes. The supervisory feedback layer monitors configuration execution data in real time (such as actual returns and risk deviations), periodically evaluates the large model's decision-making effectiveness, and automatically triggers the feedback process when data deviations exceed thresholds or evaluations fail to meet standards. The large model fine-tuning module updates the training dataset based on the feedback data, dynamically adjusts model parameters (such as optimizing weights to capture long-term trends), and achieves iterative model performance through incremental fine-tuning. This layer allows the large model to break free from the limitations of "static rigidity," ensuring the adaptability of the configuration strategy through continuous optimization and supporting the long-term stable operation of the system.

[0100] Triggering conditions: The deviation between actual and expected returns exceeds the threshold, risk indicators exceed the warning line, strategy and execution are inconsistent, and the periodic evaluation cycle has arrived.

[0101] Data flow: Feedback data flows back to the core layer of the large model, triggering incremental fine-tuning and forming a complete overall solution.

[0102] Monitoring indicators and deviation thresholds: ① Actual return deviation: threshold ±2%; ② Risk deviation: Threshold ±10%; ③ Configuration execution deviation: threshold ±3; ④ Compliance deviation: 0% (zero tolerance).

[0103] Evaluation metrics: accuracy of return forecast, risk coverage, excess return, maximum drawdown, and compliance execution rate.

[0104] Evaluation cycles: real-time, intraday, daily, weekly, monthly, quarterly.

[0105] Large-scale model fine-tuning mechanism: Employs incremental sample updates and low-learning-rate top-level fine-tuning (learning rate 1e-5~5e-5, 1~3 rounds). Automatic optimization is performed when the model fails to capture long-term trends. a. Long-period feature weights: ×1.2~2.0; b. Short-term noise weight: ×0.5~0.8; c. Strengthen long-context attention and the weighting of annuity business rule features; Through continuous iterative optimization, the model can break free from static rigidity and maintain long-term adaptive capability.

[0106] To facilitate understanding of how this invention is implemented, an example is given below.

[0107] Suppose a corporate annuity plan has initial assets of 100 million yuan, which are invested in three asset classes: stocks, bonds, and cash. Under the traditional approach, the allocation ratio is 30% for stocks, 60% for bonds, and 10% for cash.

[0108] After using the annuity asset data processing method based on a large model provided in this embodiment of the invention: 1. Data Collection: The system collects daily stock market data, bond price fluctuation data, macroeconomic data, and the latest rules for pension asset operation annuity business.

[0109] 2. Data preprocessing: Clean the stock market data and remove outliers; perform natural language processing on the text of pension asset operation annuity business rules to extract key information, such as adjusting the upper limit of stock asset operation ratio to 40%.

[0110] 3. Model training: Train the large model using historical data and currently collected data, and adjust the model parameters.

[0111] 4. Strategy Generation: The large model generates an initial allocation strategy with a stock allocation ratio of 35%, bond allocation of 55%, and cash allocation of 10%.

[0112] 5. Dynamic Optimization: If the stock market experiences a sharp decline on a given day, triggering a threshold, the system initiates a rebalancing process. After a reassessment by the large model, the stock allocation is adjusted to 30%, bonds to 60%, and cash to 10%.

[0113] 6. Risk control and compliance review: After review, the adjusted allocation strategy complies with the regulatory requirements for the upper limit of the proportion of stocks in pension fund asset operations.

[0114] 7. Strategy Execution and Feedback: The adjusted allocation strategy will be implemented in the operation of annuity assets, and the model will be optimized based on the results of asset operation.

[0115] The beneficial effects of the annuity asset data processing method based on a large model provided in this invention include: 1. Improve market adaptability: The large model can capture market dynamics in real time and dynamically adjust asset allocation strategies, enabling annuity assets to better adapt to the market environment.

[0116] 2. Enhance data integration capabilities: Effectively integrate multi-source data, fully explore the valuable information in the data, and provide a more accurate basis for asset allocation decisions.

[0117] 3. Enhance risk warning capabilities: Monitor risk indicators in real time, identify and warn of emerging risks in a timely manner, and reduce the potential risk range of annuity asset operation.

[0118] 4. Improve decision-making efficiency: Automated data processing and configuration strategy generation reduce manual intervention, improve decision-making efficiency, and meet the high-frequency adjustment needs of annuity asset operations.

[0119] Enhance model interpretability: Enhance the interpretability of large models through techniques such as causal reasoning, making configuration strategies more compliant with regulatory requirements and business logic.

[0120] In summary, the annuity asset data processing method based on a large model provided in this embodiment of the invention includes: Multi-source heterogeneous data fusion and preprocessing technology: Integrates domain operational data (stock, bond, and fund market data), annuity business rule data (pension asset management annuity business rules, tax annuity business rules), enterprise data (financial data, operational data), and macroeconomic data (GDP, inflation rate, interest rate) to form multi-dimensional data input. Natural Language Processing (NLP) technology is used to parse unstructured data such as annuity business rule texts and news reports, extracting key information and transforming it into structured input, providing semantic-level support for model training. Format differences between different data sources are eliminated, missing values ​​are filled, and outliers are handled to ensure data quality and consistency, improving the stability of model training.

[0121] Large-scale model architecture design and domain adaptation technology: Based on pre-trained large-scale models (such as BERT and LLM), leveraging their powerful language understanding and generation capabilities to process text data. Fine-tuning techniques are used to adapt the model to the professional terminology and rules of the pension asset allocation field, and a domain knowledge base is combined to enhance the model's sensitivity to pension asset operation rules and market dynamics.

[0122] Simultaneously, feature extraction and joint modeling are performed on structured domain operational data and unstructured text data to achieve cross-modal information complementarity and improve decision-making accuracy.

[0123] This invention also provides a large-model-based annuity asset data processing device, as described in the following embodiments. Since the principle by which this device solves the problem is similar to the large-model-based annuity asset data processing method, the implementation of this device can refer to the implementation of the large-model-based annuity asset data processing method; repeated details will not be elaborated further.

[0124] Figure 4 This is a schematic diagram of the structure of a pension asset data processing device based on a large model in one embodiment of the present invention, as shown below. Figure 4 As shown, the device includes: The acquisition unit 01 is used to acquire multi-source heterogeneous data in the annuity business scenario, and preprocess the multi-source heterogeneous data to obtain standardized annuity original basic feature data. Parallel parsing unit 02 is used to extract annuity operational feature data from the original basic annuity feature data, input the annuity operational feature data into the annuity operational feature data recognition model, and obtain annuity operational feature data recognition results; it is also used to extract annuity risk feature data from the original basic annuity feature data, input the annuity risk feature data into the annuity risk assessment model, and obtain annuity quantitative risk assessment results; the annuity operational feature data recognition model is pre-trained and generated from historical annuity operational feature data and corresponding recognition result samples, and the annuity risk assessment model is pre-trained and generated from historical annuity risk feature data and corresponding quantitative risk assessment result samples. Processing unit 03 is used to input the original basic characteristic data of annuities, the identification results of the operational characteristic data of annuities, and the quantitative risk assessment results of annuities into the configuration optimization model. The configuration optimization model is based on reinforcement learning to perform multiple rounds of iterative optimization and adaptively generate annuity asset allocation ratio data.

[0125] In one embodiment, the above-mentioned large-model-based annuity asset data processing apparatus may further include: The first extraction unit is used to extract the first classification label, which serves as the global feature aggregation identifier, relative to the original basic feature data of annuities, the identification results of operational feature data of annuities, and the quantitative risk assessment results of annuities in the preset layer network of the annuity field operation feature data identification model. The second extraction unit is used to extract the second attention weight of the second classification label, which serves as the global feature aggregation identifier, relative to the identification results of the original basic feature data of annuities, the operational feature data of annuities, and the quantitative risk assessment results of annuities in the preset layer network of the annuity risk assessment model. A fusion unit is used to fuse the first attention weight and the second attention weight to generate a total attention weight matrix; The aggregation weight determination unit is used to aggregate the total attention weight matrix according to four dimensions: business rule keywords, macro indicators, domain operation data, and annuity-specific risk factors, to obtain the aggregate weight of each dimension. The weight percentage determination unit is used to aggregate weights based on each dimension, and calculate the ratio of the aggregated weight of a single dimension to the total weight of all dimensions through a preset normalization calculation method, and convert it into a percentage to obtain the weight percentage of each dimension. The visualization output unit is used to complete visualization rendering based on the weight ratio of each dimension. Each set of annuity asset allocation ratio data corresponds to a unique dimension weight distribution feature. Based on the weight distribution feature, a matching annuity allocation decision explanation text is generated to realize the visualization output of the annuity allocation decision explanation text.

[0126] In one embodiment, the above-mentioned large-model-based annuity asset data processing apparatus may further include: The pressure verification processing unit is used to send the generated annuity asset allocation ratio data back to the annuity risk assessment model to perform annuity risk pressure verification processing and detect whether the overall risk indicators of the current annuity asset allocation ratio data meet the preset safety threshold. The feedback unit is used to feed back the current annuity asset allocation ratio data to the allocation optimization model if the risk indicators do not meet the preset safety threshold. This allows for multiple rounds of iterative optimization to be performed again until the asset operation risk indicators meet the preset safety threshold, and then outputs compliant annuity asset allocation ratio data.

[0127] In one embodiment, the above-mentioned large-model-based annuity asset data processing apparatus may further include: The monitoring unit is used to monitor the actual execution results of annuity asset allocation ratio data, and to collect operational efficiency deviation data, annuity asset operation risk deviation data, and compliance execution data. The update unit is used to trigger incremental fine-tuning of the configuration optimization model when various deviation data exceed the corresponding preset threshold, update the parameters of the configuration optimization model, and regenerate annuity asset allocation ratio data adapted to changes in industry and operational risks.

[0128] In one embodiment, the configuration optimization model uses the original basic characteristic data of annuities and the identification results of operational characteristic data in the annuity field as input data to complete the inference calculation, and uses the quantitative risk assessment results of annuities as optimization constraints to complete multiple rounds of iterative optimization based on the Actor-Critic reinforcement learning algorithm.

[0129] In one embodiment, the multi-source heterogeneous data includes structured data and unstructured data; the structured data includes annuity account data, market transaction data, and macroeconomic indicators, and the unstructured data includes business rule text, industry dynamic information text, and corporate financial report text; the preprocessing of the multi-source heterogeneous data includes: performing normalization, one-hot encoding, and missing value interpolation on the structured data, and performing redundant information removal, word segmentation, and semantic embedding vectorization on the unstructured data.

[0130] Based on the aforementioned inventive concept, such as Figure 5 As shown, the present invention also proposes a computer device 500, including a memory 510, a processor 520, and a computer program 530 stored in the memory 510 and executable on the processor 520. When the processor 520 executes the computer program 530, it implements the aforementioned annuity asset data processing method based on a large model.

[0131] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for processing annuity asset data based on a large model.

[0132] This invention also provides a computer program product, which includes a computer program that, when executed by a processor, implements the above-described method for processing annuity asset data based on a large model.

[0133] In this embodiment of the invention, the annuity asset data processing scheme based on a large model innovatively adopts a dual-branch parallel parsing architecture, relying on standardized original annuity basic feature data. It simultaneously performs two independent large model operations on the original basic data: one path extracts annuity domain operational features and inputs them into a pre-trained annuity domain operational feature data recognition large model to complete the quantitative identification of annuity domain operations; the other path extracts annuity risk features and inputs them into a pre-trained annuity risk assessment model to achieve quantitative assessment and analysis of asset risk. This embodiment of the invention differs from the traditional single-path serial, shallow mining static processing mode. By using a dual-model parallel independent parsing approach, it deeply mines the hidden correlation features between multi-source data, significantly improving the information utilization rate and feature mining completeness of the original data, effectively solving the problems of one-sided feature mining, information loss, and analytical bias in traditional techniques. Simultaneously, relying on the correlation reasoning capabilities of the aforementioned two dedicated large models—the annuity domain operational feature data recognition large model and the annuity risk assessment model—this invention performs holistic linkage analysis of multiple types of related asset data, overcoming the limitations of traditional static, isolated, and single-dimensional data processing. Building upon this foundation, this invention further integrates three types of multidimensional data: original basic annuity characteristic data, annuity operational characteristic data identification results, and quantitative risk assessment results. These data are input into a configuration optimization model and, relying on reinforcement learning algorithms, achieve multi-round iterative optimization dynamic calculations. This continuously and adaptively optimizes the data processing logic and calculation results, completely breaking through the outdated processing mode of traditional static solidification and one-time output. It significantly improves the calculation accuracy, multidimensional correlation analysis capabilities, and adaptability to complex market scenarios in annuity asset data analysis, providing highly accurate and stable data support for the subsequent refined and intelligent allocation and output of annuity assets.

[0134] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0135] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0136] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0137] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0138] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for processing annuity asset data based on a large model, characterized in that, include: Collect multi-source heterogeneous data in the context of annuity business, preprocess the multi-source heterogeneous data, and obtain standardized annuity raw basic feature data; Annuity operational characteristic data is extracted from the original basic annuity characteristic data, and the annuity operational characteristic data is input into the annuity operational characteristic data recognition model to obtain the annuity operational characteristic data recognition result; annuity risk characteristic data is extracted from the original basic annuity characteristic data, and the annuity risk characteristic data is input into the annuity risk assessment model to obtain the annuity quantitative risk assessment result; the annuity operational characteristic data recognition model is pre-trained and generated from historical annuity operational characteristic data and corresponding recognition result samples, and the annuity risk assessment model is pre-trained and generated from historical annuity risk characteristic data and corresponding quantitative risk assessment result samples. The original basic characteristic data of annuities, the identification results of operational characteristic data of annuities, and the quantitative risk assessment results of annuities are input into the configuration optimization model. The configuration optimization model is based on reinforcement learning and performs multiple rounds of iterative optimization to adaptively generate annuity asset allocation ratio data.

2. The method as described in claim 1, characterized in that, Also includes: In the preset-layer network of the annuity field operation feature data identification model, the first classification label, which serves as the global feature aggregation identifier, is the first attention weight relative to the original basic feature data of annuities, the annuity field operation feature data identification results, and the annuity quantitative risk assessment results. In the pre-defined layer network of the annuity risk assessment model, the second classification label, which serves as the global feature aggregation identifier, is used as the second attention weight relative to the original basic feature data of annuities, the identification results of operational feature data in the annuity field, and the quantitative risk assessment results of annuities. The first attention weight and the second attention weight are combined to generate a total attention weight matrix; Based on four dimensions—business rule keywords, macro indicators, domain operation data, and annuity-specific risk factors—the total attention weight matrix is ​​aggregated to obtain the aggregated weights for each dimension. Based on the aggregated weights of each dimension, the aggregated weight of a single dimension is calculated as a ratio to the total weight of all dimensions using a preset normalization method, and then converted into a percentage to obtain the weight percentage of each dimension. Visualization is achieved based on the weight ratio of each dimension. Each set of annuity asset allocation ratio data corresponds to a unique dimension weight distribution feature. Based on the weight distribution feature, a matching annuity allocation decision explanation text is generated, realizing the visual output of the annuity allocation decision explanation text.

3. The method as described in claim 1, characterized in that, Also includes: The generated annuity asset allocation ratio data is sent back to the annuity risk assessment model for annuity risk pressure verification, and the overall risk indicators of the current annuity asset allocation ratio data are checked to see if they meet the preset safety threshold. If the risk indicators do not meet the preset safety threshold, the current annuity asset allocation ratio data will be fed back to the allocation optimization model, and multiple rounds of iterative optimization will be performed again until the asset operation risk indicators meet the preset safety threshold, and compliant annuity asset allocation ratio data will be output.

4. The method as described in claim 1, characterized in that, Also includes: Monitor the actual implementation results of annuity asset allocation ratio data, and collect data on operational efficiency deviations, annuity asset operation risk deviations, and compliance execution data. When various deviation data exceed the corresponding preset thresholds, the incremental fine-tuning of the configuration optimization model is triggered, the parameters of the configuration optimization model are updated, and the annuity asset allocation ratio data adapted to changes in industry and operational risks is regenerated.

5. The method as described in claim 1, characterized in that, The configuration optimization model uses the original basic characteristic data of annuities and the identification results of operational characteristic data in the field of annuities as input data to complete the inference calculation, and uses the quantitative risk assessment results of annuities as the optimization constraint condition to complete multiple rounds of iterative optimization based on the Actor-Critic reinforcement learning algorithm.

6. The method as described in claim 1, characterized in that, The multi-source heterogeneous data includes structured data and unstructured data; the structured data includes annuity account data, market transaction data, and macroeconomic indicators, while the unstructured data includes business rule text, industry dynamic information text, and corporate financial report text. The preprocessing of the multi-source heterogeneous data includes: performing normalization, one-hot encoding and missing value interpolation on structured data, and performing redundant information removal, word segmentation and semantic embedding vectorization on unstructured data.

7. A data processing device for pension assets based on a large model, characterized in that, include: The data acquisition unit is used to collect multi-source heterogeneous data in the annuity business scenario, and to preprocess the multi-source heterogeneous data to obtain standardized annuity raw basic feature data. The parallel parsing unit is used to extract annuity operational feature data from the original basic annuity feature data, input the annuity operational feature data into the annuity operational feature data recognition model, and obtain annuity operational feature data recognition results; it also extracts annuity risk feature data from the original basic annuity feature data, inputs the annuity risk feature data into the annuity risk assessment model, and obtains annuity quantitative risk assessment results; the annuity operational feature data recognition model is pre-trained and generated from historical annuity operational feature data and corresponding recognition result samples, and the annuity risk assessment model is pre-trained and generated from historical annuity risk feature data and corresponding quantitative risk assessment result samples. The processing unit is used to input the original basic characteristic data of annuities, the identification results of the operational characteristic data of annuities, and the quantitative risk assessment results of annuities into the configuration optimization model. The configuration optimization model is based on reinforcement learning to perform multiple rounds of iterative optimization and adaptively generate annuity asset allocation ratio data.

8. 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 method of any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method of any one of claims 1 to 6.

10. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the method of any one of claims 1 to 6.

Images