Intelligent financial risk early warning method, system, device and storage medium

Abstract

The application discloses a kind of intelligent financial risk early warning method, system, equipment and storage medium.The intelligent financial risk early warning method is by building the multimodal time series dataset containing numerical features and text features, the first analysis sub-path and the second analysis sub-path of parallel execution, realizes the double capture of financial text content context semantics and key risk signals;Through time alignment and feature aggregation, the multimodal features generated at different times from different sources are unified to the standardized time unit, and a unified risk feature vector is generated;By using time series cross-validation strategy to train the model, integrating multiple machine learning algorithms, and introducing a fixed effect logistic regression model, the interference of time trend and group difference is effectively controlled, and the advantages of different algorithms are integrated.The application can discover the potential risks of financial entities earlier, more comprehensively and more accurately, and provide decision support for investors and regulatory authorities.

Description

Technical Field

[0001] This application belongs to the field of information technology, and in particular relates to an intelligent financial risk early warning method, system, device and storage medium. Background Technology

[0002] Existing financial risk early warning and assessment methods suffer from the following main shortcomings: First, they rely on a limited range of data types. Traditional rule-based methods or some artificial intelligence models primarily depend on structured financial or credit data, making it difficult to fully utilize unstructured information such as news, announcements, social media, and images. Second, their risk identification capabilities are limited. The reliance on rules or the limited model makes it difficult to accurately assess complex events and potential hidden risks. Third, they lack dynamic adaptability. Existing systems are slow to respond to emerging risks and market changes, unable to update predictive models and response strategies in real time. Fourth, the predictive results lack interpretability. Users struggle to understand the causes and basis of risks, thus limiting their application value in practical decision-making and strategy formulation. For example, in patent CN110009475B, the risk audit and monitoring method, device, computer equipment, and storage medium mainly compare and identify anomalies in financial data by setting audit rules. While this can improve the efficiency of compliance checks to some extent, it relies too heavily on fixed manual rules, lacks adaptive learning capabilities, cannot react quickly to newly emerging financial risk events, and struggles to handle unstructured data. Therefore, this method is insufficient in terms of the comprehensiveness and real-time nature of risk identification. Summary of the Invention

[0003] The purpose of this application is to provide an intelligent financial risk early warning method, system, device and storage medium that can at least partially solve the shortcomings of existing financial risk early warning and assessment methods.

[0004] According to a first aspect of this application, this application provides an intelligent financial risk early warning method, which includes the following steps:

[0005] S1: Acquire and standardize financial entity data from heterogeneous data sources to construct a multimodal time series dataset containing both numerical and textual features;

[0006] Among them, numerical features include at least one of financial factors, governance factors and / or market factors, and textual features include financial text content related to financial entities;

[0007] S2: Based on textual features, the first and second analysis sub-paths are processed in parallel.

[0008] The first analysis sub-path is configured to: use a pre-trained large language model to analyze the deep semantics of financial text content, and generate context-based sentiment scores and financial fraud risk level labels.

[0009] The second analysis sub-path is configured as follows: based on the financial fraud seed terminology library, a dynamic risk terminology library is constructed through a weak supervision expansion mechanism, and the weighted frequency of risk keywords in financial text content is calculated based on the dynamic risk terminology library to generate financial fraud risk factors;

[0010] S3: Align and aggregate the emotional score, financial fraud risk level label, and financial fraud risk factor with the numerical features according to the preset time frequency to generate a unified risk feature vector of the financial entity in each time unit.

[0011] S4: A time series cross-validation strategy is used to divide a unified risk feature vector. Based on the divided unified risk feature vector, a fixed-effects logistic regression model is trained, and multiple machine learning classification algorithms are integrated to output the risk probability prediction results of financial entities in the future time window.

[0012] In some embodiments, step S1 includes:

[0013] Acquire financial entity data from heterogeneous data sources;

[0014] Financial entity data undergoes data cleaning, numerical conversion, and time-series alignment to be standardized to a quarterly frequency.

[0015] Each processed financial entity data entry is labeled with its corresponding financial entity code and timestamp, thus forming a multimodal time-series dataset.

[0016] In some embodiments, the deep semantic analysis using a pre-trained large language model in step S2 includes:

[0017] Construct prompt word templates and input financial text content into the large language model;

[0018] This enables large language models to perform sentiment analysis based on prompt word templates and output sentiment scores.

[0019] And / or, enable the large language model to perform financial fraud risk classification based on prompt word templates and output financial fraud risk level labels.

[0020] In some embodiments, constructing a dynamic risk lexicon through a weakly supervised expansion mechanism includes:

[0021] Use a pre-defined seed word library for financial fraud as the initial word set;

[0022] The initial word set is semantically expanded using a word vector model to obtain candidate words with similar meanings;

[0023] Candidate words are incorporated into the initial word set to form an expanded dynamic risk word library.

[0024] In some embodiments, the step S2 of calculating the weighted frequency of risk keywords in financial text content to generate a financial fraud risk factor includes:

[0025] Based on a dynamic risk terminology database, financial text content is segmented and filtered for stop words, and the frequency of occurrence of risk keywords is statistically analyzed.

[0026] The semantic strength index is calculated based on the frequency of occurrence.

[0027] The information concentration index is calculated based on the inverse impact factor, which is the ratio of the total number of all financial entities within a preset time window to the number of financial entities mentioned in the financial text.

[0028] Multiply the semantic strength index by the information concentration index to generate a weighted risk signal value for each piece of financial text.

[0029] The system aggregates the weighted risk signal values ​​of all financial text content within a preset time window to generate a financial fraud risk factor.

[0030] In some embodiments, the preset time frequency in step S3 includes:

[0031] Financial entity data generated between January 1 and April 30 each year will be included in the previous year's financial reporting cycle.

[0032] Financial entity data generated between May 1 and December 31 each year will be included in the financial reporting cycle of that year.

[0033] In some embodiments, step S4, which involves using a time-series cross-validation strategy to partition a unified risk feature vector and then training a fixed-effects logistic regression model based on the partitioned unified risk feature vector, includes the following steps:

[0034] The unified risk feature vector is divided into a training set and a test set according to the time order, so that the time window of the training set is earlier than the time window of the corresponding test set.

[0035] A logistic regression model was constructed, with the risk probability prediction result of the future time window as the dependent variable and the financial fraud risk factor, financial factor and market factor in the unified risk feature vector as independent variables.

[0036] Introducing year-fixed and peer-group-fixed effects into the logistic regression model to control for time-series fluctuations and group differences;

[0037] The maximum likelihood estimation method is used to train a logistic regression model with fixed effects of year and peer group based on the training set, and a fixed effects logistic regression model is obtained.

[0038] In some embodiments, the step of integrating multiple machine learning classification algorithms for model training in step S4 includes:

[0039] The initial risk probability prediction results generated by the fixed effects logistic regression model are used as meta-features and concatenated with the unified risk feature vector;

[0040] The concatenated feature vectors are input into the ensemble learning framework, enabling the ensemble learning framework to run multiple base classifier algorithms in parallel.

[0041] By integrating the outputs of multiple base classifier algorithms through voting or weighted averaging mechanisms, risk probability prediction results are generated.

[0042] Among them, the various base classifier algorithms are selected from at least two of the linear model, tree model, support vector machine, neural network and Boosting algorithm.

[0043] In some embodiments, the method further includes the following after step S4:

[0044] Monitor and acquire newly generated financial entity data. When newly generated financial entity data within a preset time window is captured, automatically execute an incremental learning process to update the integrated learning framework and dynamic risk vocabulary. Use the updated integrated learning framework and dynamic risk vocabulary to generate new risk probability prediction results.

[0045] The incremental learning process includes:

[0046] Based on the newly generated financial entity data, the base classifier in the ensemble learning framework is incrementally trained to adjust its weights; based on the financial text content of the newly generated financial entity data, the dynamic risk lexicon is expanded and updated.

[0047] In some embodiments, the method further includes the following after step S4:

[0048] Based on the risk probability prediction results, a pre-constructed financial domain knowledge graph is used to perform strategy reasoning to generate basic strategies. Then, a reinforcement learning model based on historical feedback is used to dynamically optimize the basic strategies, thereby generating market-adaptive strategies for different user types.

[0049] A second aspect of this application provides an intelligent financial risk early warning system, the system comprising:

[0050] The data fusion module is configured to acquire and standardize financial entity data from heterogeneous data sources, and construct a multimodal time-series dataset containing both numerical and textual features.

[0051] Among them, numerical features include at least one of financial factors, governance factors and / or market factors, and textual features include financial text content related to financial entities;

[0052] The dual-path risk signal extraction module is configured to process the first analysis sub-path and the second analysis sub-path in parallel based on textual features.

[0053] The first analysis sub-path is configured to: use a pre-trained large language model to analyze the deep semantics of financial text content, and generate context-based sentiment scores and financial fraud risk level labels.

[0054] The second analysis sub-path is configured as follows: based on the financial fraud seed terminology library, a dynamic risk terminology library is constructed through a weak supervision expansion mechanism, and the weighted frequency of risk keywords in financial text content is calculated based on the dynamic risk terminology library to generate financial fraud risk factors;

[0055] The feature alignment and aggregation module is configured to align and aggregate sentiment scores, financial fraud risk level labels, and financial fraud risk factors with numerical features according to a preset time frequency, generating a unified risk feature vector for financial entities in each time unit.

[0056] The risk probability prediction module is configured to use a time series cross-validation strategy to divide a unified risk feature vector, train a fixed-effects logistic regression model based on the divided unified risk feature vector, and integrate multiple machine learning classification algorithms to output the risk probability prediction results of financial entities in the future time window.

[0057] A third aspect of this application provides an electronic device including a memory and a processor coupled to each other, the processor being used to execute program instructions stored in the memory to implement the intelligent financial risk early warning method of any of the above.

[0058] A fourth aspect of this application provides a computer-readable storage medium having program instructions stored thereon, which, when executed by a processor, implement the intelligent financial risk early warning method described above.

[0059] The intelligent financial risk early warning method of this application constructs a multimodal time-series dataset containing both numerical and textual features. This combines traditional quantitative data with textual information implying deep risks, overcoming the limitations of a single data source and improving the comprehensiveness and depth of risk identification. Through parallel execution of the first and second analysis sub-paths, it achieves dual capture of the contextual semantics of textual information and key risk signals, enhancing the detection capability and robustness against hidden risks. By time-series alignment and feature aggregation, multimodal features from different sources and at different times are unified onto standardized time units, generating a unified risk feature vector, ensuring the timeliness and consistency of risk features. By employing a time-series cross-validation strategy to train the model, it strictly adheres to the chronological order of financial entity data, avoiding future information leakage and making the model evaluation results more realistic and reliable. By integrating multiple machine learning algorithms and introducing a fixed-effects logistic regression model, it effectively controls the interference of time trends and group differences, combining the advantages of different algorithms. This application can discover potential risks of financial entities earlier, more comprehensively, and more accurately, especially hidden risks such as financial fraud, providing decision support for investors and regulatory agencies. Attached Figure Description

[0060] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0061] Figure 1 This is a flowchart illustrating an embodiment of the intelligent financial risk early warning method of this application.

[0062] Figure 2 yes Figure 1 A flowchart of step S1.

[0063] Figure 3 yes Figure 1 A flowchart illustrating the process of deep semantic analysis using a pre-trained large language model in step S2.

[0064] Figure 4 yes Figure 1 A flowchart illustrating the process of constructing a dynamic risk lexicon through a weakly supervised expansion mechanism in step S2.

[0065] Figure 5 yes Figure 1 A flowchart illustrating the process of calculating the weighted frequency of risk keywords in financial text content in step S2 to generate financial fraud risk factors.

[0066] Figure 6 yes Figure 1 The flowchart of step S4, which uses time series cross-validation to divide a unified risk feature vector, and then trains a fixed-effects logistic regression model based on the divided unified risk feature vector, is shown.

[0067] Figure 7 Figure 1 A schematic diagram of the process of integrating multiple machine learning classification algorithms for model training in step S4.

[0068] Figure 8 yes Figure 1 A flowchart illustrating the process after step S4.

[0069] Figure 9 This is a partial flowchart of another embodiment of the intelligent financial risk early warning method of this application.

[0070] Figure 10 This is a schematic diagram of the structure of an embodiment of the intelligent financial risk early warning system of this application.

[0071] Figure 11 This is a schematic diagram of the hardware structure of an embodiment of the electronic device of this application.

[0072] Figure 12 This is a schematic diagram of the structure of an embodiment of the computer-readable storage medium of this application. Detailed Implementation

[0073] To better understand the above-mentioned objectives, features, and advantages of this disclosure, the solutions disclosed herein will be further described below. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0074] Numerous specific details are set forth in the following description in order to provide a full understanding of this disclosure, but this disclosure may also be implemented in other ways different from those described herein; obviously, the embodiments in the specification are only some, and not all, of the embodiments of this disclosure.

[0075] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the term "comprising" or any other variations thereof is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0076] The intelligent financial risk early warning method, system, device, and storage medium provided in this application will be described in detail below with reference to the accompanying drawings and through specific embodiments and application scenarios.

[0077] Please see Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the intelligent financial risk early warning method of this application. The intelligent financial risk early warning method includes:

[0078] S1: Acquire and standardize financial entity data from heterogeneous data sources to construct a multimodal time series dataset containing both numerical and textual features;

[0079] Among them, numerical features include at least one of financial factors, governance factors and / or market factors, and textual features include financial text content related to financial entities;

[0080] Specifically, financial entities are the target of this analysis, and typically include: listed companies, unlisted companies (especially those planning to go public or issue bonds), financial institutions (such as banks, insurance companies, securities companies, etc.), and other legal entities active in the capital market.

[0081] Heterogeneous data sources are the sources of financial entity data, including exchange announcements, financial news websites, corporate information disclosure platforms, social media opinion platforms, etc. Financial entity data can be selectively obtained periodically from these heterogeneous data sources. Financial entity data refers to the collection of all data that can be used to assess the risk status of a specific financial entity.

[0082] Financial entity data typically includes both numerical and textual features. Numerical features are usually structured data, including at least one of financial factors, governance factors, and / or market factors. Financial factors refer to data directly obtained from financial statements, such as debt-to-equity ratio, net profit, revenue growth rate, and cash flow; governance factors reflect data on corporate governance structure, such as shareholding concentration, board size, proportion of independent directors, and executive compensation; market factors reflect data on a company's performance in the capital market, such as stock returns, volatility, turnover rate, and market capitalization.

[0083] Text-based features refer to unstructured data, including financial text content related to financial entities. This financial text content includes official disclosures, such as company annual reports, semi-annual reports, prospectuses, and announcements of major events; financial news and commentary, such as reports and analyses about the company from major financial media outlets and news websites; and analyst research reports, such as in-depth research reports on the company published by securities firms. The formats of this financial text content include text, images, video, and audio.

[0084] Financial entity data all contain a key attribute: timestamps, such as financial reporting period, news release date, etc., which is the basis for subsequent time-series alignment.

[0085] Therefore, financial entity data is a massive dataset with both numerical and textual features and time-series information that can be used to analyze corporate risk; in other words, it is a multimodal time-series dataset.

[0086] In one specific implementation, such as Figure 2 As shown, this step specifically includes:

[0087] S101: Obtain financial entity data from heterogeneous data sources;

[0088] S102: Perform data cleaning, numerical conversion, and time-series alignment on financial entity data to unify it to a quarterly frequency;

[0089] Data cleaning and numerical transformation are used to clean up erroneous data in financial entity data, and then time-series alignment is used to unify all financial entity data to a quarterly frequency.

[0090] S103: Label each processed financial entity data with its corresponding financial entity code and timestamp to form a multimodal time series dataset.

[0091] Specifically, the financial entity code is an individual index of a financial entity. Each financial entity corresponds to a unique financial entity code, which can be defined by the securities code, the unified social credit code, or the internal identification number. This financial entity code ensures that subsequent models, such as the big oracle model, word vector model, and fixed-effects logistic regression model, can track the dynamic changes of financial fraud risk factors over time at the financial entity level. Through fixed-effects modeling, it is possible to effectively control the temporal fluctuations and group differences between different financial entities caused by differences in size, industry structure, or governance mechanisms, thereby focusing on the dynamic changes of the risk indicators themselves.

[0092] A timestamp is a time-period index, serving as the time dimension for subsequent models. Timestamps can be set as annual, quarterly, or sliding window formats depending on the research objectives. For example, when using an annual time unit, timestamps correspond to the years 2020, 2021, 2022, ...; when using a quarterly unit, timestamps correspond to 2020Q1, 2020Q2, ... Each timestamp represents a closed time window, used to statistically analyze all news samples, language features, and model prediction indicators within that period. The introduction of timestamps can capture the dynamic trends of risk indicators, constructing a future-oriented prediction structure, thereby supporting time series modeling and logistic regression prediction.

[0093] S2: Based on textual features, the first and second analysis sub-paths are processed in parallel.

[0094] The first analysis sub-path is configured to: use a pre-trained large language model to analyze the deep semantics of financial text content, and generate context-based sentiment scores and financial fraud risk level labels.

[0095] The second analysis sub-path is configured as follows: based on the financial fraud seed terminology library, a dynamic risk terminology library is constructed through a weak supervision expansion mechanism, and the weighted frequency of risk keywords in financial text content is calculated based on the dynamic risk terminology library to generate financial fraud risk factors;

[0096] Specifically, this step employs a dual analysis mechanism to corroborate each other. It leverages the reasoning capabilities of a large language model to uncover potential risks in complex contexts (such as management intentions and event correlations), while a statistically based dynamic risk lexicon ensures the stability and interpretability of risk signal extraction, reducing the probability of misjudgment and missed judgment.

[0097] In one specific implementation, such as Figure 3 As shown, step S2, which involves performing deep semantic analysis using a pre-trained large language model, includes:

[0098] S201: Construct prompt word templates and input financial text content into the large language model;

[0099] The prompt word template contains keywords related to financial fraud, such as "financial fraud," "false disclosure," "fraud," and "audit opinion," to guide the large language model in accurate risk identification.

[0100] Large language models can perform deep semantic understanding of financial texts. Compared with traditional rule-based or dictionary-based methods, large language models have the ability to perform contextual reasoning and financial context recognition, and can accurately capture keywords and key phrases related to risk events such as financial fraud, illegal disclosure, regulatory penalties, and market doubts in financial news and stock forum comments. Optionally, the large language model is the Qwen3-32B model.

[0101] S202: Enables large language models to perform sentiment analysis based on prompt word templates and output sentiment scores;

[0102] Guided by the aforementioned prompt word template, the large language model performs sentiment analysis and outputs a sentiment score between 0 and 1 for each piece of financial text, labeled as ES, where 0 represents extremely negative, 0.5 represents neutral, and 1 represents extremely positive.

[0103] Based on sentiment scores, the system automatically identifies and marks financial texts with extremely negative sentiment scores. Subsequently, according to a preset time frequency, it counts the number of financial texts marked as extremely negative within the same time unit.

[0104] Finally, this statistical result is aligned with a complete and continuous time series to ensure that each time unit within the preset historical time window has corresponding statistical data on the quantity of extremely negative financial text content, providing structured input for the subsequent generation of a unified risk feature vector.

[0105] S203: And / or, enable the large language model to perform financial fraud risk classification based on prompt word templates and output financial fraud risk level labels.

[0106] Guided by prompt word templates, the large language model performs a financial fraud risk classification task and outputs financial fraud risk level labels, including no risk, low risk, or high risk level labels, providing highly intelligent support for financial fraud risk early warning.

[0107] In one specific implementation, such as Figure 4 As shown, the steps for constructing a dynamic risk lexicon through a weakly supervised expansion mechanism include:

[0108] S21: Use the preset financial fraud seed word library as the initial word set;

[0109] The financial fraud seed term database combines the functions of both a financial seed term database and a fraud seed term database, forming a comprehensive financial fraud seed term database.

[0110] S22: Use word vector models to semantically expand the initial word set and obtain candidate words with similar meanings;

[0111] The word vector model can be chosen as the Word2Vec model. Using the Word2Vec model, the specified seed words in the initial word set are semantically expanded to obtain candidate words that are semantically similar to them in the vector space. The Word2Vec model can capture implicit synonyms and derivative terms in the context. It belongs to semi-automatic lexicon construction and is often used in natural language processing literature for expanding sentiment dictionaries and financial risk dictionaries. For example, "faking" can be expanded to "inflating" or "falsely reporting".

[0112] S23: Incorporate candidate words into the initial word set to form an expanded dynamic risk word library.

[0113] In one specific implementation, such as Figure 5 As shown, step S2, which calculates the weighted frequency of risk keywords in financial text content to generate a financial fraud risk factor, includes:

[0114] S24: Based on a dynamic risk thesaurus, perform word segmentation and stop word filtering on financial text content, and count the frequency of occurrence of risk keywords;

[0115] Specifically, for processing financial text content of a single news article, Jieba segmentation is used to segment the financial text content into words, and stop word filtering is applied to the financial text content to filter out stop words such as "company", "enterprise", "group", etc., retaining potential keywords of 2-4 characters in length for keyword matching.

[0116] Based on a dynamic risk term database, it is determined whether the financial text of the news article contains risky keywords, such as "profit," "report," "cost," and "violation." If they appear, the news article is considered to be fraud-related, and the frequency of the risky keywords is counted. Otherwise, they may be ignored or given very low weight.

[0117] S25: Calculate the semantic strength index based on the frequency of occurrence;

[0118] Based on the frequency of occurrence of risk keywords, the total number of times all risk keywords appear in the news article is calculated. The higher the total number of occurrences, the stronger the discussion of financial fraud in the news article is, and the stronger the risk signal. Therefore, it can be used as a semantic strength indicator and marked as ATF.

[0119] S26: Calculate the information concentration index based on the inverse impact factor, where the inverse impact factor is the ratio of the total number of all financial entities within a preset time window to the number of financial entities mentioned in the financial text.

[0120] The inverse impact factor is defined as the ratio of the total number of all financial entities within a preset time window to the number of financial entities mentioned in the financial text content, i.e., the total number of all financial entities within the preset time window / the number of financial entities mentioned in the financial text content. If the news article involves only a few companies, the inverse impact factor is larger, reflecting a high concentration of information, and therefore it can be used as an information concentration indicator and labeled as IIF.

[0121] S27: Multiply the semantic strength index and the information concentration index to generate a weighted risk signal value for each piece of financial text content;

[0122] Specifically, the following calculation formula can be used:

[0123] weight=ATF×IIF×(1 + c×sgn(d-ES)),

[0124] Where weight represents the weighted risk signal value, c and d are adjustment coefficients, set to 5% and 0.5 respectively, ES represents the sentiment score, and sgn() represents the sign function, that is, if the value of d-ES is greater than 0, it takes 1, if the value of d-ES is less than 0, it takes -1, and if it is equal to 0, it takes 0.

[0125] S28: Aggregate the weighted risk signal values ​​of all financial text content of financial entities within a preset time window to generate a financial fraud risk factor.

[0126] FRI i,t =weight_sum i,t / N i,t

[0127] Among them FRI i,t N represents the risk factor for financial fraud. i,t This represents the number of articles in which the statistical company was mentioned in the news within a preset time window, where i represents the i-th financial entity and t represents the t-th timestamp.

[0128] S3: Align and aggregate the emotional score, financial fraud risk level label, and financial fraud risk factor with the numerical features according to the preset time frequency to generate a unified risk feature vector of the financial entity in each time unit.

[0129] Specifically, the preset time frequency adopts the accounting year mapping rule. Chinese listed companies usually disclose their annual reports before April 30 of the following year. Therefore, the financial entity data generated between January 1 and April 30 of each year are included in the previous year's financial reporting cycle.

[0130] Financial entity data generated between May 1 and December 31 each year will be included in the financial reporting cycle of that year.

[0131] Optionally, the emotional score ranges from [0, 1], the financial fraud risk level labels are no risk (represented by the value 0), low risk (represented by the value 1), and high risk (represented by the value 2), and the financial fraud risk factor is a quantified risk intensity value. These are all structured data.

[0132] Numerical features come from multimodal time-series datasets and are a type of structured data.

[0133] All of the above structured data are associated with specific financial entity company codes and timestamps.

[0134] The emotional score, financial fraud risk level label, and financial fraud risk factor are aligned and aggregated with the numerical features according to a preset time frequency. For example, the time unit for alignment and aggregation is "quarter". The alignment rules are as follows:

[0135] For sentiment scores, financial fraud risk level labels, and financial fraud risk factors, the system assigns them to the corresponding calendar quarter, such as 2023Q1, based on their original release timestamp.

[0136] For numerical features, since they are usually released quarterly, the quarterly data from their official reports are used directly.

[0137] To achieve uniform alignment, the system maintains a master timeline. For financial entities that do not have official financial data within a preset quarter, a forward padding strategy is adopted, which uses data from the most recent available quarter to ensure the integrity of the feature vectors.

[0138] Within each time unit, such as Q1 2023, various types of data for each financial entity are aggregated:

[0139] Sentiment score aggregation: Calculate the arithmetic mean of the sentiment scores of all relevant texts within the quarter;

[0140] Financial fraud risk level tag aggregation: Statistics on the distribution of different financial fraud risk level tags within the quarter, and calculation of the proportion of high-risk tags;

[0141] Financial fraud risk factor aggregation: Calculate the arithmetic mean of the financial fraud risk factor values ​​of all financial text content within the quarter to represent the average risk intensity of the quarter.

[0142] Numerical feature processing: The time-aligned numerical features are combined with the sentiment scores, financial fraud risk level labels, and financial fraud risk factors obtained from the above aggregation to generate a unified risk feature vector.

[0143] Once the aggregation is complete, the system will generate a unified risk feature vector for each financial entity each quarter. This vector is structured and can be directly called by machine learning models.

[0144] S4: A time series cross-validation strategy is used to divide a unified risk feature vector. Based on the divided unified risk feature vector, a fixed-effects logistic regression model is trained, and multiple machine learning classification algorithms are integrated to output the risk probability prediction results of financial entities in the future time window.

[0145] Specifically, in one embodiment, such as Figure 6 As shown, step S4, which involves using a time-series cross-validation strategy to partition a unified risk feature vector and then training a fixed-effects logistic regression model based on this partitioned unified risk feature vector, includes the following steps:

[0146] S401: Divide the unified risk feature vector into a training set and a test set according to the time order, so that the time window of the training set is earlier than the time window of its corresponding test set.

[0147] Specifically, the unified risk feature vectors are sorted in descending order of time, and an initial training time window is set. For example, if the unified risk feature vectors from 2001 to 2010 are used as the initial training set, then the unified risk feature vectors from 2011 are used as the first test set.

[0148] The model is trained on the initial training set and then used for prediction and performance evaluation on the corresponding test set.

[0149] Subsequently, a rolling window validation was performed, incorporating the unified risk feature vector from 2011 into the training set. That is, the unified risk feature vectors from 2001 to 2011 were used as the training set, and the unified risk feature vector from 2012 was used as the new test set. The training and testing process was repeated.

[0150] This process continues until all available annual data is traversed. This method ensures that the time windows for all uniform risk feature vectors used for training are earlier than the corresponding test set time windows, thereby preventing future information leaks.

[0151] S402: Construct a logistic regression model with the risk probability prediction result of the future time window as the dependent variable and the financial fraud risk factor, financial factor and market factor in the unified risk feature vector as independent variables.

[0152] Specifically, a logistic regression model is constructed, with the predicted risk probability of a future time window as the dependent variable, indicating whether a financial entity will experience risk in the next year.

[0153] The financial fraud risk factor, financial factor, and market factor in the unified risk feature vector are used as independent variables. Among them, the financial fraud risk factor is the core predictor variable, which indicates the risk intensity in the current year; the financial factor and market factor are control variables to control the influence of other factors.

[0154] S403: Introduce year-fixed effects and peer-group fixed effects into the logistic regression model to control for time-series fluctuations and group differences;

[0155] Introducing year-fixed effects and peer-group fixed effects into the logistic regression model. The year-fixed effect involves generating a dummy variable FE for each sample year. year This is used to capture the systematic impact of macroeconomic or time trends on the risk of financial entities; peer group fixed effects divide financial entities into different "peer groups" based on their industry classification (such as the primary industry codes of the China Securities Regulatory Commission), and generate a dummy variable FE for each peer group with a sufficient sample size. peer Encode a sufficient number of peer groups in the sample to control for unobserved heterogeneity within the same industry or market category. Groups with low sample size are uniformly classified as "other". The total number of dummy variables can be limited to avoid overfitting or the curse of dimensionality.

[0156] S404: Using the maximum likelihood estimation method, a logistic regression model with fixed effects of year and peer group is trained based on the training set to obtain a fixed effects logistic regression model.

[0157] The mathematical form of the fixed-effects logistic regression model constructed based on steps S401-S403 is as follows:

[0158]

[0159] Among them, Y i,t+1 This indicates whether the i-th company has experienced financial fraud or a high-risk event in the period t+1 (1 indicates that financial fraud or a high-risk event has occurred, and 0 indicates that no financial fraud or a high-risk event has occurred).

[0160] FRI i,t Indicates the risk factor for financial fraud; X i,t For the enterprise, the control variable at timestamp t; FE year For year-fixed effects; FE peer This is a peer group fixed effect;

[0161] β0 is the intercept term, which represents the baseline log odds of financial fraud or high-risk events occurring when all independent variables are zero and the fixed effects are at the baseline level.

[0162] β1 measures the core predictor variable FRI i,t The marginal impact on the probability of financial fraud or high-risk events is as follows: if β1>0, it indicates that the higher the financial fraud risk factor, the greater the probability of financial fraud or high-risk events occurring in the future; conversely, if β1<0, the higher the financial fraud risk factor, the lower the probability of financial fraud or high-risk events occurring in the future.

[0163] β2 is a control variable coefficient used to measure the impact of other financial entity characteristics on the probability of financial fraud or high-risk events, thus making β1 more reflective of FRI. i,t The net effect.

[0164] The parameter β of the above model is estimated using the maximum likelihood estimation method. The model is then trained based on the training set. After training, a fixed-effects logistic regression model with defined parameters for a specific training time window is obtained. This model is used to calculate the risk probability prediction results for the test set samples of financial entities in the future time window.

[0165] To further improve the accuracy and robustness of financial entity risk prediction, in one specific implementation, after training the fixed-effects logistic regression model in the above steps, as follows: Figure 7 As shown, it also includes integrating multiple machine learning classification algorithms for model training, with the following steps:

[0166] S41: Use the initial risk probability prediction results generated by the fixed effects logistic regression model as meta-features and concatenate them with the unified risk feature vector;

[0167] The fixed-effects logistic regression model obtained through steps S401-S404, and the risk probability prediction results generated based on this model, are called the initial risk probability prediction results. As a new feature, it is also called the meta-feature, which represents the primary model's judgment on the risk of the sample.

[0168] The meta-feature is concatenated with the original unified risk feature vector to form a new, enhanced feature vector with increased dimensionality. This enhanced feature vector contains both the original multimodal risk information and the prediction priors of the primary model.

[0169] S42: Input the concatenated feature vector into the ensemble learning framework, enabling the ensemble learning framework to run multiple base classifier algorithms in parallel;

[0170] The ensemble learning framework can run multiple base classifier algorithms in parallel. These base classifier algorithms are selected from at least two of linear models, tree models, support vector machines, neural networks, and Boosting algorithms, with the two base classifier algorithms preferably having complementary performance.

[0171] Linear model algorithms can include L1 regularized logistic regression, L2 regularized logistic regression, elastic network logistic regression, or stochastic gradient descent logistic regression. Tree model algorithms can include decision trees, random forests, extreme random trees, gradient boosting decision trees, or histogram gradient boosting. Support vector machine algorithms can include radial basis function kernel-based support vector machines or linear support vector machines. Neural network algorithms can include multilayer perceptrons. Booting algorithms can include XGBoost, LightGBM, or CatBoost.

[0172] The concatenated feature vector, i.e. the enhanced feature vector, is used as a new input to the ensemble learning framework to train all selected base classifiers on the training set.

[0173] S43: Integrate the outputs of multiple base classifier algorithms through voting or weighted averaging mechanisms to generate risk probability prediction results.

[0174] After the base classifier is trained, it is used to predict the test set samples. One of the following two ensemble mechanisms can be used:

[0175] Voting mechanism: Each base classifier outputs a discrete risk probability prediction result (such as "high risk" or "low risk"). The final result is determined by voting from all classifiers, and the result with the most votes is the final risk probability prediction result.

[0176] Weighted averaging mechanism: The training set is further divided chronologically within its time window, using a portion of the time window as the training set. For example, using the unified risk feature vector from 2001 to 2010 as the training set, and selecting data from 2008 to 2010 as the validation set. The performance of each base classifier is evaluated on the validation set, and its score is used as the basis for weight allocation. The score indicates how well the model can distinguish between "high-risk" and "low-risk" financial entities. The higher the score, the better the discriminative ability of the base classifier, and the higher its performance. The higher the weight of the base classifier with better performance, the greater its weight.

[0177] Each base classifier outputs a continuous risk probability prediction for the test set samples. The final risk probability prediction is the weighted average of the probabilities output by each base classifier.

[0178] Through the series of steps described above, the ensemble learning framework outputs the final risk probability prediction result obtained through the ensemble mechanism. This result demonstrates higher accuracy and robustness compared to predictions from a single model.

[0179] Existing financial risk early warning methods are slow to respond to emerging risks and market changes, and cannot update predictive models and response strategies in real time. In addition to the steps described above, this application's embodiments, to improve the dynamic adaptability of the intelligent financial risk early warning method, include, for example... Figure 8 As shown, after step S4, the following is also included:

[0180] S5: Monitor and acquire newly generated financial entity data. When newly generated financial entity data within a preset time window is captured, automatically execute the incremental learning process to update the integrated learning framework and dynamic risk lexicon.

[0181] S6: Use the updated ensemble learning framework and dynamic risk lexicon to generate new risk probability prediction results;

[0182] The incremental learning process includes:

[0183] Based on the newly generated financial entity data, the base classifier in the ensemble learning framework is incrementally trained; based on the financial text content of the newly generated financial entity data, the dynamic risk lexicon is expanded and updated.

[0184] Specifically, the system deploys a background daemon process to continuously monitor designated financial data sources, including news APIs, exchange announcement streams, corporate information platforms, and social media, to obtain newly generated financial entity data. This newly generated financial entity data also includes both numerical and textual features.

[0185] The system is pre-defined with a specific time window (e.g., a quarter) as the trigger time window for incremental learning. Once the system has captured and accumulated new data for a complete time window (e.g., the second quarter of 2025), the incremental learning process is automatically triggered.

[0186] The newly generated financial entity data within this complete time window are processed according to steps S1-S3 to generate a new unified risk feature vector. Then, the base classifier in the ensemble learning framework is incrementally trained based on the new unified risk feature vector.

[0187] If the ensemble framework uses a voting mechanism: the purpose of incremental training is to directly improve the classification accuracy of each base classifier, thereby improving the consensus quality of the final voting results as a whole. Under this mechanism, each classifier usually maintains equal voting rights. If the ensemble framework uses a weighted average mechanism: the performance of all base classifiers is re-evaluated on the updated validation set, and their weights in the ensemble framework are recalculated and reassigned based on the new performance.

[0188] For newly generated financial entity data, deep semantic analysis is performed on the financial text content in the financial entity data using a pre-set large language model to identify and extract keywords and key phrases related to new events, new terms, and new policies.

[0189] High-quality new risk terms identified by the large language model are added to the dynamic risk lexicon. Simultaneously, semantic extension models such as Word2Vec word vectors are continued to be used. Based on the existing financial fraud seed lexicon and high-quality new risk terms, vectorization training is performed on a new initial word set to capture the latest semantically relevant words, further expanding the coverage of the dynamic risk lexicon.

[0190] Through the above process, the dynamic risk lexicon was expanded, thereby improving the semantic recognition capability for new risk phenomena.

[0191] After completing the incremental learning process, the updated integrated learning framework and dynamic risk lexicon are used to conduct a risk assessment of the current financial entities and generate new risk probability prediction results.

[0192] Through this periodic incremental learning process, the method in this embodiment ensures that the fixed-effects logistic regression model, the integrated learning framework, and the dynamic risk lexicon can keep up with market dynamics, enabling continuous, accurate, and forward-looking early warning of financial risks, thereby achieving dynamic optimization and strategy adjustment of the intelligent financial risk early warning method.

[0193] The intelligent financial risk early warning method of this application embodiment achieves dual capture of textual information context semantics and key risk signals through multimodal data fusion and parallel execution of the first and second analysis sub-paths, enhancing the detection capability and robustness of hidden risks. Through rigorous time series processing and an advanced machine learning algorithm framework, it constructs a comprehensive, high-precision, highly robust, and self-evolving intelligent financial risk early warning method, effectively solving many pain points of traditional methods in terms of comprehensiveness, accuracy, and practicality.

[0194] like Figure 9 As shown, another embodiment of the intelligent financial risk early warning method of this application, in addition to including the methods of the above embodiments, includes, as... Figure 9 As shown, after step S4, the following is also included:

[0195] S7: Based on the risk probability prediction results, strategy reasoning is performed in conjunction with a pre-built financial domain knowledge graph to generate basic strategies. Then, a reinforcement learning model based on historical feedback is used to dynamically optimize the basic strategies, thereby generating market-adaptive strategies for different user types.

[0196] Specifically, based on the risk probability prediction results, strategy reasoning is performed by combining a pre-built financial domain knowledge graph, and the strategy is dynamically optimized using a reinforcement learning model based on historical feedback to generate market-adaptive strategies for different user types.

[0197] Based on the risk probability prediction results and using the financial entity code as a query condition, a pre-built financial domain knowledge graph is accessed. This financial domain knowledge graph stores the relationships between financial entities (such as upstream and downstream of the industrial chain, equity relationships), historical risk events and response cases, financial product information, and regulatory policies.

[0198] Based on the interconnected information of a knowledge graph in the financial field, the system executes reasoning logic to generate actionable market adaptation strategies. For example, for high-risk financial entities, it automatically links them to their upstream and downstream companies, guarantors, etc., suggesting avoidance, reducing positions, increasing hedging positions, and prompting in-depth due diligence; for medium-risk financial entities, it suggests close monitoring, setting stop-loss lines, diversifying investments, and pushing relevant industry competitive analysis reports; for low-risk financial entities, it can recommend them as potential investment opportunities, provided they match the user's risk appetite. It can also generate differentiated and personalized market adaptation strategies based on different user types. For example, it can generate concise "buy / hold / sell" suggestions for individual investors, and detailed reports including complex strategies such as position adjustments and derivative hedging for institutional investors.

[0199] The system records each generated strategy and the user's actual execution (or simulated execution results), and uses the actual gains / losses of the strategy as historical feedback. The reinforcement learning model uses the strategy and its feedback as training data to continuously optimize its strategy selection model. Its goal is to learn which strategies can bring the highest long-term returns under specific market environments and risk signals, thereby achieving dynamic adaptation to market changes.

[0200] Once generated, market adaptation strategies can be presented in various ways. For example, a market adaptation strategy report can integrate risk probability predictions, strategy recommendations, and explanations based on a financial knowledge graph. Alternatively, it can be presented as interactive trend charts or risk distribution heatmaps, visually demonstrating risk trends and industry risk clustering. Furthermore, the market adaptation strategy report and visualizations are output through system interfaces and can be displayed on different terminals, including a PC-based management backend and a mobile app, ensuring users can receive and execute strategies anytime, anywhere.

[0201] Through the above implementation methods, the intelligent financial risk early warning method of this application embodiment completes the closed loop from risk prediction to strategy execution, providing users with intelligent, personalized and executable action guidelines, and greatly enhancing the practical value of the method of this application embodiment.

[0202] Please see Figure 10 , Figure 10 This is a schematic diagram of an embodiment of the intelligent financial risk early warning system of this application. The intelligent risk early warning system includes a data fusion module 201, a dual-path risk signal extraction module 202, a feature alignment and aggregation module 203, and a risk probability prediction module 204. These modules are directly or indirectly connected in communication.

[0203] The data fusion module 201 is configured to acquire and standardize financial entity data from heterogeneous data sources, and construct a multimodal time series dataset containing numerical and textual features.

[0204] Among them, numerical features include at least one of financial factors, governance factors and / or market factors, and textual features include financial text content related to financial entities;

[0205] The dual-path risk signal extraction module 202 is configured to process the first analysis sub-path and the second analysis sub-path in parallel based on textual features.

[0206] The first analysis sub-path is configured to: use a pre-trained large language model to analyze the deep semantics of financial text content, and generate context-based sentiment scores and financial fraud risk level labels.

[0207] The second analysis sub-path is configured as follows: based on the financial fraud seed terminology library, a dynamic risk terminology library is constructed through a weak supervision expansion mechanism, and the weighted frequency of risk keywords in financial text content is calculated based on the dynamic risk terminology library to generate financial fraud risk factors.

[0208] The feature alignment and aggregation module 203 is configured to align and aggregate sentiment scores, financial fraud risk level labels, and financial fraud risk factors with numerical features according to a preset time frequency, thereby generating a unified risk feature vector for the financial entity in each time unit.

[0209] The risk probability prediction module 204 is configured to use a time series cross-validation strategy to divide a unified risk feature vector, train a fixed effects logistic regression model based on the divided unified risk feature vector, and integrate multiple machine learning classification algorithms to output the risk probability prediction results of financial entities in the future time window.

[0210] It should be noted that those skilled in the art will understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0211] Please see Figure 11 , Figure 11 This is a schematic diagram of the hardware structure of an embodiment of the electronic device of this application. The device includes a memory and a processor coupled to each other. The processor is used to execute program instructions stored in the memory to implement the corresponding process as described in the foregoing method embodiment. In a specific implementation scenario, the electronic device 30 may include, but is not limited to, a microcomputer or a server. In addition, the electronic device 30 may also include mobile devices such as laptops and tablets, which are not limited here.

[0212] The electronic device may include a processor 301 and a memory 302 storing program instructions.

[0213] When processor 301 executes the program, it implements the steps in any of the above method embodiments.

[0214] For example, the program can be divided into one or more modules / units, one or more of which are stored in memory 302 and executed by processor 301 to complete this application. The one or more modules / units can be a series of program instruction segments capable of performing a specific function, which describe the execution process of the program in the device.

[0215] Specifically, the processor 301 may include a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of this application.

[0216] Memory 302 may include mass storage for data or instructions. For example, and not limitingly, memory 302 may include a hard disk drive (HDD), floppy disk drive, flash memory, optical disk, magneto-optical disk, magnetic tape, or Universal Serial Bus (USB) drive, or a combination of two or more of these. Where appropriate, memory 302 may include removable or non-removable (or fixed) media. Where appropriate, memory 302 may be internal or external to the integrated gateway disaster recovery device. In a particular embodiment, memory 302 is non-volatile solid-state memory.

[0217] Memory may include read-only memory (ROM), random access memory (RAM), disk storage media devices, optical storage media devices, flash memory devices, and electrical, optical, or other physical / tangible memory storage devices. Therefore, typically, memory includes one or more tangible (non-transitory) readable storage media (e.g., memory devices) encoded with software including computer-executable instructions, and when the software is executed (e.g., by one or more processors), it is operable to perform the operations described with reference to the methods according to one aspect of this disclosure.

[0218] The processor 301 implements any of the methods described above by reading and executing program instructions stored in the memory 302.

[0219] In one example, the electronic device may also include a communication interface 303 and a bus 310. The processor 301, memory 302, and communication interface 303 are connected via the bus 310 and communicate with each other.

[0220] The communication interface 303 is mainly used to realize communication between various modules, devices, units and / or equipment in the embodiments of this application.

[0221] Bus 310 includes hardware, software, or both, that couples components of an online data traffic metering device together. For example, and not limitingly, the bus may include an Accelerated Graphics Port (AGP) or other graphics bus, an Enhanced Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), HyperTransport (HT) interconnect, an Industry Standard Architecture (ISA) bus, an Infinite Bandwidth Interconnect, a Low Pin Count (LPC) bus, a memory bus, a Microchannel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-X) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local (VLB) bus, or other suitable buses, or combinations of two or more of these. Where appropriate, bus 310 may include one or more buses. Although specific buses are described and illustrated in embodiments of this application, any suitable bus or interconnect is contemplated herein.

[0222] In addition, the process in conjunction with the foregoing method embodiments, such as Figure 12 As shown, this application embodiment can be implemented using a computer-readable storage medium 40. The storage medium 40 stores program instructions 401; these program instructions 401 are executed by a processor to implement any step in the above method embodiment.

[0223] This application also provides a chip, which includes a processor and a communication interface. The communication interface and the processor are coupled. The processor is used to run programs or instructions to implement the various processes of the above method embodiments and achieve the same technical effect. To avoid repetition, it will not be described again here.

[0224] It should be understood that the chip mentioned in the embodiments of this application may also be referred to as a system-on-a-chip, system chip, chip system, or system-on-a-chip, etc.

[0225] This application provides a computer program product, which is stored in a storage medium and executed by at least one processor to implement the various processes of the above method embodiments and achieve the same technical effects. To avoid repetition, it will not be described again here.

[0226] It should be clarified that this application is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of this application is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of this application.

[0227] The functional modules shown in the above block diagram can be implemented as hardware, software, firmware, or a combination thereof. When implemented in hardware, they can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this application are programs or code segments used to perform the required tasks. Programs or code segments can be stored on machine-readable media or transmitted over a transmission medium or communication link via data signals carried on a carrier wave. "Machine-readable media" can include any medium capable of storing or transmitting information. Examples of machine-readable media include electronic circuits, semiconductor memory devices, ROM, flash memory, erasable ROM (EROM), floppy disks, CD-ROMs, optical disks, hard disks, fiber optic media, radio frequency (RF) links, etc. Code segments can be downloaded via computer grids such as the Internet, intranets, etc.

[0228] The aspects of this disclosure have been described above with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and program products according to embodiments of this disclosure. It should be understood that each block in 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 program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to create a machine such that these instructions, executable via the processor of the computer or other programmable data processing apparatus, enable the implementation of the functions / actions specified in one or more blocks of the flowchart illustrations and / or block diagrams. Such a processor can be, but is not limited to, a general-purpose processor, a special-purpose processor, a special application processor, or a field-programmable logic circuit. It is also understood that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can also be implemented by special-purpose hardware performing the specified functions or actions, or can be implemented by a combination of special-purpose hardware and computer instructions.

[0229] The above are merely specific embodiments of this application. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, modules, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. It should be understood that the protection scope of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the protection scope of this application.

Claims

1. A smart financial risk early warning method, characterized in that, The method comprises the following steps: S1: acquiring and standardizing financial entity data from heterogeneous data sources, and constructing a multi-modal time series dataset containing numerical features and text features; wherein the numerical features include at least one of financial factors, governance factors and / or market factors, and the text features include financial and economic text content related to the financial entity; S2: processing a first analysis sub-path and a second analysis sub-path in parallel based on the text features; wherein the first analysis sub-path is configured to analyze the deep semantics of the financial and economic text content using a pre-trained large language model, generate a context-based sentiment score and a financial fraud risk level label; the second analysis sub-path is configured to construct a dynamic risk word library based on a financial fraud seed word library through a weakly supervised expansion mechanism, and calculate the weighted frequency of risk keywords in the financial and economic text content based on the dynamic risk word library, and generate a financial fraud risk factor; S3: aligning and aggregating the sentiment score, the financial fraud risk level label and the financial fraud risk factor with the numerical features at a preset time frequency, and generating a unified risk feature vector of the financial entity at each time unit; S4: dividing the unified risk feature vector using a time series cross-validation strategy, training a fixed effect logistic regression model based on the divided unified risk feature vector, and integrating multiple machine learning classification algorithms to output risk probability prediction results of the financial entity in a future time window.

2. The method of claim 1, wherein, The step S1 comprises: acquiring the financial entity data from heterogeneous data sources; performing data cleaning, numerical conversion and time series alignment processing on the financial entity data to unify to a quarterly frequency; labeling the corresponding financial entity code and timestamp for each processed financial entity data, thereby forming the multi-modal time series dataset.

3. The method of claim 1, wherein, The deep semantic analysis using a pre-trained large language model in the step S2 comprises: building a prompt word template and inputting the financial and economic text content into the large language model; making the large language model perform sentiment analysis based on the prompt word template and output the sentiment score; and / or, making the large language model perform financial fraud risk classification based on the prompt word template and output the financial fraud risk level label.

4. The method of claim 1, wherein, The construction of the dynamic risk word library through a weakly supervised expansion mechanism comprises: using a preset financial fraud seed word library as an initial word set; using a word vector model to perform semantic expansion on the initial word set to obtain candidate words with similar semantics; integrating the candidate words into the initial word set to form an expanded dynamic risk word library.

5. The method of claim 4, wherein, The step of calculating the weighted frequency of risk keywords in the financial and economic text content to generate a financial fraud risk factor in the step S2 comprises: based on the dynamic risk word library, performing word segmentation and stop word filtering processing on the financial and economic text content, and counting the occurrence frequency of the risk keywords; calculating a semantic intensity indicator according to the occurrence frequency; calculating an information concentration indicator according to an inverse influence factor, wherein the inverse influence factor is a ratio of a total number of all the financial entities within a preset time window to a number of the financial entities mentioned in the financial text content; multiplying the semantic intensity indicator and the information concentration indicator to generate a weighted risk signal value of each piece of financial text content; aggregating the weighted risk signal values of the financial entity in all the financial text content within the preset time window to generate the financial fraud risk factor.

6. The method of claim 1, wherein, The preset time frequency in step S3 comprises: financial entity data generated from January 1 to April 30 of each year is attributed to the financial report cycle of the previous year; financial entity data generated from May 1 to December 31 of each year is attributed to the financial report cycle of the current year.

7. The method of claim 1, wherein, The step of dividing the uniform risk feature vector based on the time series cross-validation strategy in step S4 comprises: dividing the uniform risk feature vector into a training set and a test set in chronological order, so that the time window of the training set is earlier than the time window of the corresponding test set; constructing a logistic regression model, taking the risk probability prediction result of the future time window as the dependent variable, and taking the financial fraud risk factor, the financial factor and the market factor in the uniform risk feature vector as the independent variable; introducing year fixed effects and peer group fixed effects into the logistic regression model to control time series fluctuations and group differences; training the logistic regression model with year fixed effects and peer group fixed effects based on the training set by using the maximum likelihood estimation method to obtain the fixed effects logistic regression model.

8. The method of claim 7, wherein, The step of training the model by integrating multiple machine learning classification algorithms in step S4 comprises: concatenating the initial risk probability prediction result generated by the fixed effects logistic regression model as meta-features with the uniform risk feature vector; inputting the concatenated feature vector into an ensemble learning framework, and making the ensemble learning framework run multiple base classifier algorithms in parallel; integrating the outputs of the multiple base classifier algorithms by voting or weighted average mechanism to generate the risk probability prediction result; wherein the multiple base classifier algorithms are selected from at least two of linear models, tree models, support vector machines, neural networks and Boosting algorithms.

9. The method of claim 8, wherein, After step S4, it further comprises: monitoring and acquiring newly generated financial entity data, and when the newly generated financial entity data of a preset time window is captured, automatically performing an incremental learning process to update the ensemble learning framework and the dynamic risk vocabulary; using the updated ensemble learning framework and the dynamic risk vocabulary to generate a new risk probability prediction result; wherein the incremental learning process comprises: incrementally training the base classifiers in the ensemble learning framework based on the newly generated financial entity data to adjust the weights of the base classifiers; extending and updating the dynamic risk vocabulary based on the financial text content of the newly generated financial entity data.

10. The method according to any one of claims 1 to 9, characterized in that, After step S4, it further comprises: Based on the risk probability prediction result, strategy reasoning is performed in combination with a pre-constructed financial field knowledge graph to generate a basic strategy, and a reinforcement learning model based on historical feedback is used to dynamically optimize the basic strategy, thereby generating market adaptation strategies for different user types.

11. An intelligent financial risk early warning system, characterized in that, The system comprises: A data fusion module configured to acquire and standardize financial entity data from heterogeneous data sources, and to construct a multi-modal time series dataset containing numerical features and text features; wherein the numerical features include at least one of financial factors, governance factors, and / or market factors, and the text features include financial text content related to the financial entity; A dual-path risk signal extraction module configured to process a first analysis sub-path and a second analysis sub-path in parallel based on the text features; wherein the first analysis sub-path is configured to analyze the deep semantics of the financial text content using a pre-trained large language model, and to generate a context-based sentiment score and a financial fraud risk level label; the second analysis sub-path is configured to construct a dynamic risk word library based on a financial fraud seed word library through a weak supervision expansion mechanism, and to calculate the weighted frequency of risk keywords in the financial text content based on the dynamic risk word library, and to generate a financial fraud risk factor; A feature alignment and aggregation module configured to align and aggregate the sentiment score, the financial fraud risk level label, and the financial fraud risk factor with the numerical features at a preset time frequency to generate a unified risk feature vector of the financial entity at each time unit; A risk probability prediction module configured to divide the unified risk feature vector using a time series cross-validation strategy, train a fixed effect logistic regression model based on the divided unified risk feature vector, and integrate multiple machine learning classification algorithms to output a risk probability prediction result of the financial entity in a future time window.

12. An electronic device, comprising: The program instructions are executed by the processor to implement the intelligent financial risk early warning method of any one of claims 1 to 10.

13. A computer-readable storage medium having stored thereon program instructions, wherein the program instructions are executable by a computer for causing the computer to perform the method according to any one of claims 1 to 12. The program instructions are executed by the processor to implement the intelligent financial risk early warning method of any one of claims 1 to 10.

Images