Enterprise financial risk prediction method and system based on multi-source feature fusion
By integrating social media comment data, MD&A data, and financial ratio data, the MFIFN model is constructed, which solves the problem of inaccurate financial risk prediction caused by single-perspective analysis in existing technologies, and achieves more accurate corporate financial risk prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-26
- Publication Date
- 2026-04-14
AI Technical Summary
In existing technologies, corporate financial risk prediction models based on financial data analyze textual data from only a single perspective, resulting in inaccurate analysis results that cannot be used as decisive information.
A multi-source feature fusion method is adopted. The FinBERT algorithm is used to analyze social networking site comment data, the LLM-BERT model is used to analyze MD&A data, and financial ratio data is combined to construct the MFIFN model to calculate the ST probability of enterprises. The social networking site comment data, MD&A data and financial ratio data are integrated, and the Focal loss algorithm is used for binary classification prediction.
It improves the accuracy of corporate financial risk prediction and the reliability of judgment. By integrating multi-source data, it retains key risk characteristics as much as possible and reduces the error of analysis results.
Smart Images
Figure CN121860424A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of financial management technology, and in particular to a method and system for predicting corporate financial risks based on multi-source feature fusion. Background Technology
[0002] A company's financial risk is a gradually worsening process. Timely and effective forecasting helps managers intervene in advance and formulate appropriate management measures. Financial risk arises from both internal and external factors. Internally, a company's financial data is used to assess its financial condition and predict financial risk. Externally, increased macroeconomic volatility, fierce market competition, and growing consumer and investor concern about product quality and operational performance make risk prediction based solely on financial data analysis more challenging. Therefore, risk management must consider both internal and external information.
[0003] A company's financial risk is influenced by a variety of factors. Numerical financial data only covers quantitative information about the company and does not include qualitative information. Using only financial data to predict future financial risks is incomplete. Therefore, more and more researchers are beginning to combine textual financial data to provide incremental information for financial risk prediction. Current text analysis models mainly include sentiment analysis and semantic analysis. However, these methods only analyze the readability, semantics, or sentiment of the text from a single perspective. They do not consider the information expressed by the entire text from multiple perspectives such as attributes, viewpoints, and sentiments. This results in inaccurate analysis results, which can only be used as reference data and cannot be used as decisive information. Summary of the Invention
[0004] To address the shortcomings of existing technologies, this invention provides a method and system for predicting corporate financial risks based on multi-source feature fusion. This method solves the problem that existing analytical models analyze textual financial data from a single perspective, resulting in inaccurate analysis results.
[0005] According to an embodiment of the present invention, a method for predicting corporate financial risk based on multi-source feature fusion includes: We acquire social media comment data and use the FinBERT algorithm to perform semantic sentiment analysis on the social media comment data to obtain the first statistical feature set. Obtain MD&A data, use the LLM-BERT model to perform semantic sentiment analysis on the MD&A data, and obtain the second statistical feature set and semantic feature set; Obtain financial data and calculate financial ratios based on the financial data; The MFIFN model is constructed based on the first statistical feature set, the second statistical feature set, the semantic feature set, and financial ratio data. The MFIFN model is used to calculate the ST probability of a company and to judge the financial risk of the company based on the ST probability.
[0006] Preferably, the method for using the FinBERT algorithm to perform semantic sentiment analysis on social networking site comment data to obtain the first statistical feature set includes: The FinBERT algorithm was used to classify the sentiment of social media comment data, resulting in positive sentiment group, neutral sentiment group and negative sentiment group; Based on the data volume of the positive sentiment group and the negative sentiment group, calculate the first relative positivity and the first overall positivity between the positive sentiment group and the negative sentiment group; Calculate the percentage of data in the positive sentiment group, neutral sentiment group, and negative sentiment group relative to the total data volume; The data volume and corresponding data proportion of the positive sentiment group, neutral sentiment group, and negative sentiment group, the total data volume of social networking site comments, the first relative positivity, and the first overall positivity are combined into the first statistical feature set.
[0007] Preferably, the method for performing semantic sentiment analysis on MD&A data using the LLM-BERT model to obtain the second statistical feature set and semantic feature set includes: The LLM model was used to extract sentiment features from the MD&A data, resulting in positive, neutral, and negative triples. The BERT algorithm was used to perform semantic analysis on positive triples, neutral triples, and negative triples respectively, resulting in positive semantic feature sets, neutral semantic feature sets, and negative semantic feature sets. We obtain the semantic feature set by weighting all features in the positive semantic feature set, the neutral semantic feature set, and the negative semantic feature set. Based on the amount of data in the positive and negative triplets, calculate the second relative positivity and the second overall positivity between the positive triplets and the triadic sentiment groups; Calculate the percentage of data in positive triples, neutral triples, and negative triples relative to the total data volume; The data volume and corresponding data proportion in the positive, neutral and negative triplet groups, the total data volume of MD&A data, the second relative positivity and the second overall positivity are combined to form the second statistical feature set.
[0008] Preferably, the formula for calculating the first relative positivity is as follows: in, and These represent the data volume in the positive sentiment group and the negative sentiment group within the time period t.
[0009] Preferably, the method for calculating the ST probability of a firm using the MFIFN model includes: Feature extraction is performed on the first statistical feature set, the second statistical feature set, the semantic feature set, and the financial ratio data respectively. All extracted features are then concatenated and fused to obtain the fused features. Based on the fusion features, the Focal loss algorithm is used to calculate the ST probability of the enterprise.
[0010] Preferably, the method for extracting features from the first statistical feature set, the second statistical feature set, the semantic feature set, and the financial ratio data, and then concatenating and fusing all the extracted features to obtain the fused features includes: Fully connected layers are used to extract features from the first statistical feature set, the second statistical feature set, and the financial ratio, respectively, to obtain the corresponding first shallow feature, second shallow feature, and third shallow feature. The LSTM model is used to extract features from the semantic feature set to obtain temporal features; A dense layer is used to extract features from the first shallow features, the second shallow features, the third shallow features, and the temporal features to obtain the corresponding deep features. Then, all deep features are concatenated and a linear attention mechanism is used to extract features to obtain the first attention feature. The second attention feature is obtained by using a linear attention mechanism to extract features from the temporal features. The first attention feature and the second attention feature are concatenated to obtain the fused feature.
[0011] Preferably, when using the LSTM model to extract features from the semantic feature set, multiple time-step temporal features are obtained. The temporal features extracted at the last time step are then imported into a dense layer for feature extraction to obtain the corresponding deep features.
[0012] Preferably, the temporal features extracted at each time step are concatenated, and then a linear attention mechanism is used for feature extraction to obtain the second attention feature.
[0013] On the other hand, according to embodiments of the present invention, a corporate financial risk prediction system based on multi-source feature fusion is also provided. This system uses the aforementioned corporate financial risk prediction method based on multi-source feature fusion, and includes: The data acquisition module is used to acquire social media comment data, MD&A data, and financial data. The data analysis module is used to perform semantic sentiment analysis on social networking site comment data using the FinBERT algorithm to obtain a first statistical feature set, to perform semantic sentiment analysis on MD&A data using the LLM-BERT model to obtain a second statistical feature set and a semantic feature set, and to calculate financial ratio data based on financial data. A model building module, which is used to build an MFIFN model; The prediction module is used to calculate the ST probability of an enterprise using the MFIFN model based on a first statistical feature set, a second statistical feature set, a semantic feature set, and financial ratio data, and to judge the financial risk of the enterprise based on the ST probability.
[0014] Compared with the prior art, the present invention has the following beneficial effects: This invention uses three types of textual data—financial ratios, social media comment data, and MD&A data—to predict corporate financial risk. The textual features of these three types of data are transformed into statistical features, and the semantic features of the MD&A data are separately fused with all statistical features to minimize the loss of key corporate risk-related features and improve model accuracy. Then, by constructing an MFIFN model to fuse the three types of data, the Focal loss algorithm is used for binary classification prediction of corporate risk, further improving the accuracy of the judgment. Attached Figure Description
[0015] Figure 1 This is a diagram illustrating the enterprise financial risk prediction method according to an embodiment of the present invention.
[0016] Figure 2 This is a semantic sentiment analysis architecture diagram for MD&A data in an embodiment of the present invention.
[0017] Figure 3 This is a diagram of the multi-source data feature fusion architecture according to an embodiment of the present invention. Detailed Implementation
[0018] The technical solutions of the present invention will be further described below with reference to the accompanying drawings and embodiments.
[0019] like Figure 1 As shown in the figure, this invention proposes a method for predicting corporate financial risk based on multi-source feature fusion, including: We acquire social media comment data and use the FinBERT algorithm to perform semantic sentiment analysis on the social media comment data to obtain the first statistical feature set. Financial media website communities contain a large amount of posts and comments related to companies, typically edited and published by investors, media outlets, businesses, or other stakeholders. These websites aggregate diverse user perspectives, helping users gain a more comprehensive understanding of company operations and market dynamics. Furthermore, the exchange of different viewpoints generates new insights. Therefore, business managers can leverage these websites to obtain a wealth of information to support their decision-making.
[0020] This invention obtains social media comment data from some highly active financial websites and uses web crawler technology to automatically collect comment data from listed companies, providing a data foundation for subsequent text analysis, sentiment judgment, and financial risk prediction.
[0021] In the data preprocessing stage, the crawled social media comment data is first cleaned and deduplicated. Data cleaning mainly involves removing posts irrelevant to the corresponding company, including advertisements, other stock-related posts, and entertainment news. Data deduplication mainly includes two aspects: firstly, during the crawling process, the same data may be crawled multiple times due to page data updates, repeated page loads, or changes in crawler link parameters; secondly, for posts by the same user on the same event with similar titles, posts are merged to remove duplicates.
[0022] In addition, in order to obtain the main semantic information of the text, the text is further processed. First, punctuation marks, numbers and non-textual content are deleted, as these are usually not meaningful and the main numbers are already reflected in the financial data. Second, the present invention uses a word segmentation tool, namely jieba, to segment Chinese text and count word frequencies, and deletes common stop words and words with a statistical frequency of less than 2% and more than 2% in all collected datasets.
[0023] To further extract text features, sentiment analysis is a commonly used method, which can obtain sentiment classification through dictionary-based or machine learning-based methods. To improve the efficiency of sentiment analysis for text in social media, this invention uses a pre-trained BERT model from the financial domain, namely FinBERT. FinBERT is pre-trained on three financial domain datasets and exhibits significant advantages in processing Chinese financial text. Through algorithm optimization and training on specific financial corpora, FinBERT significantly improves its ability to capture key information and semantic relationships in text. This characteristic enables FinBERT to provide more accurate results in tasks such as classification, retrieval, and topic modeling. Furthermore, FinBERT demonstrates excellent scalability and flexibility. It can adapt to financial text datasets of different sizes, meaning that FinBERT remains efficient and accurate whether processing small research datasets or large market monitoring corpora.
[0024] The FinBERT algorithm performs sentiment classification and statistics on preprocessed social networking site comment data, which can obtain the number of posts for each time period (year), the number of posts in the positive sentiment group, the number of posts in the neutral sentiment group, the number of posts in the negative sentiment group, the number of reads of all posts, and the corresponding number of comments. In this invention, the number of posts is referred to as the data volume.
[0025] Considering the increasing number of posts with the development of the internet, this invention adds two types of features. The first type is a percentage feature, namely the percentage of positive posts, the percentage of neutral posts, and the percentage of negative posts, which represent the proportion of posts in the corresponding sentiment group to the total data volume. The specific formula is as follows: in, They represent positive, neutral, and negative, respectively. This represents the number of posts published within the time period t. This represents the total number of posts within the time period t.
[0026] The second category is the relative proportion of positive and negative posts, calculated using the following formula: First overall enthusiasm: First relative positivity: The value ranges between -1 and 1, reflecting the overall positivity of the posts, while It not only takes into account relative positivity, but also reflects the relative number of positive posts.
[0027] In summary, the above data features are combined into the first statistical feature set. See Table 1 for details.
[0028] Table 1 Statistical characteristics of social networking site comment data Obtain MD&A data, use the LLM-BERT model to perform semantic sentiment analysis on the MD&A data, and obtain the second statistical feature set and semantic feature set; An annual report is a formal report submitted by a company or organization at the end of an accounting year, summarizing and disclosing its operating and financial status. It is also information that listed companies must disclose to the outside world every year. This invention first crawls the annual reports of target companies from the official websites of stock exchanges, which are usually in PDF format. Secondly, the data is preprocessed, converted to text format, and the Management Discussion and Analysis (MD&A) section is extracted to obtain the company's annual MD&A data.
[0029] In order to extract information from MD&A reports that may influence future market fluctuations or company financial risks, this invention not only examines the company's outlook analysis but also seeks out information of interest to external investors, such as management's expectations for the future market or their level of interest in a particular technology. However, this section of the text is lengthy, typically several thousand or even tens of thousands of words, containing key information but also many irrelevant or irrelevant comments that investors may not care about, which could affect the predictive performance of subsequent models.
[0030] While sentiment statistics based on MD&A can provide some incremental information, the lack of semantic information analysis results in weak generalization ability of the model and an inability to assist managers in information disclosure from a linguistic and grammatical perspective. To effectively address these issues, this invention employs a Large Language Model (LLM) and BERT text feature extraction method, comprising three parts: aspect sentiment triple extraction, semantic feature extraction, and feature aggregation, such as... Figure 2 As shown.
[0031] (1) Extraction of the emotional triad Aspect Sentiment Triplet Extraction (ASTE) is built upon aspect-based sentiment analysis (ABSA). This type of method performs fine-grained sentiment analysis by identifying specific aspects of text and their associated sentiments. ABSA typically involves two extraction tasks: aspect extraction and opinion extraction, resulting in aspect-opinion pairs, such as (real estate market, stable). Sentiment is then identified on these aspect-opinion pairs to obtain aspect-sentiment triples, i.e., (aspect, opinion, sentiment), such as (real estate market, stable, neutral). Large language models, trained on large, high-quality datasets, possess strong text semantic understanding capabilities and can more effectively grasp the core ideas of a text.
[0032] This invention aggregates aspect and opinion pairs according to sentiment categories, resulting in three types of high-quality text: positive triads, neutral triads, and negative triads, which are respectively used... , and express.
[0033] (2) Semantic feature extraction In the semantic feature extraction stage, this invention employs a pre-trained BERT model to extract semantic features from three types of text. This invention selects BERT-base-Chinese as the base model, importing the MD&A data corresponding to positive, neutral, and negative triples into the BERT model. The last layer of the BERT model is used as the output, and average pooling is performed to obtain the corresponding text features. ,in , .
[0034] (3) Feature aggregation Finally, the semantic understanding features are obtained through feature aggregation. First, all extracted triples are statistically analyzed to obtain four statistical features: the total number of triples, the number of positive triples, the number of neutral triples, and the number of negative triples. Then, using the same processing method as social media comment data, the corresponding second relative positivity and second overall positivity are calculated.
[0035] To effectively represent the semantic features of the text, a weighted summation method is used to fuse the three types of semantic features, as follows: in, This yields the final semantic feature set of MD&A.
[0036] Finally, all statistical features are combined into a second statistical feature set. As shown in Table 2.
[0037] Table 2 Statistical Characteristics of MD&A Data Obtain financial data and calculate financial ratios based on the financial data; Table 3 Financial Ratio Data Table For financial data, this invention extracts financial data from the China Stock Market and Accounting Research (CSMAR) database and calculates 31 financial ratios as shown in Table 2 and Table 3 based on the financial data.
[0038] The MFIFN model is constructed based on the first statistical feature set, the second statistical feature set, the semantic feature set, and financial ratio data. The MFIFN model is used to calculate the ST probability of a company and to judge the financial risk of the company based on the ST probability.
[0039] The model architecture of the MFIFN model constructed in this invention is as follows: Figure 3 As shown, for financial ratios, the input dimension is 1×31, i.e. The third shallow layer features are obtained by extracting feature representations using a fully connected layer, as shown in the formula: in This represents the activation function, specifically the rectified linear unit (ReLU), defined as follows: .
[0040] For the statistical features in the first statistical feature set corresponding to social networking site comment data and the first statistical feature set corresponding to MD&A data, this invention uses a fully connected layer for feature extraction to obtain the first shallow features. Second shallow features .
[0041] For the semantic feature set of MD&A data, the input dimension of its semantic features is 1×768. Given the high dimensionality and sequential nature of these semantic features, this invention employs a Long Short-Term Memory (LSTM) network to extract their latent representations. This LSTM network mainly consists of a forget gate, an input gate, and an output gate.
[0042] semantic feature matrix corresponding to semantic features The original 1×768 dimension is reshaped into a sequence of 24 time steps, with each time step corresponding to 32 dimensions as input, generating a new sequence. Shape. At time step t, the forget gate in the LSTM cell ( The calculation is as follows: in Represents the sigmoid activation function. The input gate is calculated using the following formula ( ), candidate units ( ) and unit state ( ): here , This represents the Hadamard product.
[0043] In time steps The output gate of the LSTM unit ( ) and the corresponding hidden state ( The calculation is as follows: The hidden state (and time sequence features) of the last time step. The temporal features are obtained as the output representation of the LSTM. .
[0044] Based on extracted shallow features The invention further employs dense layers for deep feature extraction; specifically, we apply a weighted matrix to each shallow feature. and bias terms Fully connected transformation: } in It is the ReLU activation function.
[0045] Then, all deep features are concatenated and a linear attention mechanism is used for feature extraction to obtain the first attention feature.
[0046] For the semantic feature set, this invention concatenates the temporal features extracted from each time step in the LSTM model, uses a linear attention mechanism to extract features, and then concatenates the first attention feature and the second attention feature to obtain the fused feature F.
[0047] Financial risk prediction can be viewed as a binary classification problem. It measures the degree of financial risk by identifying the probability that a company will fall into financial distress in the future. This invention is based on fusion features and uses the Focal loss algorithm to calculate the ST probability of a company, and judges the company's financial risk based on the ST probability.
[0048] To verify the superiority of the multi-source data fusion and single-perspective analysis of this invention, four experiments were designed to compare and verify different combinations of data features: the first group used 31 financial ratios as the baseline feature set: feature set A (Fin); the second group used feature set B (Fin+Soc) which included the baseline feature set and 11 social media comment data; the third group used feature set C (Fin+MD&A) which included the baseline feature set and 9 MD&A data; and the fourth group used feature set 4 (All) which included all baseline feature sets, 11 social media comment data, and 9 MD&A data. Commonly used logistic regression (LR), ANN, and XGBoost (XGB) were used for prediction.
[0049] Table 4 Comparison of prediction performance for each feature set Table 4 shows that feature sets B and C outperform feature set A in all classification methods, indicating that social media comments and MD&A data can provide incremental information about financial ratios, consistent with previous research findings. Furthermore, the performance difference between feature sets B and C is not significant, although set C performs better in most cases. Feature set D achieves the best results across all methods, demonstrating that combining social media comment data and MD&A data can provide more valuable information, thereby improving the effectiveness of financial risk prediction.
[0050] To verify the superiority of the proposed LLM-BERT text feature extraction method, this invention selects four commonly used and cutting-edge text feature extraction methods: topic-based analysis (LDA), BERT-based semantic model, dictionary-based sentiment extraction (SENTIMENT), and SFL-based method. Furthermore, this paper uses a large language model with strong Chinese comprehension capabilities—Tongyi Qianwen—to extract aspect sentiment triples from MD&A. Finally, the proposed deep learning-based financial risk prediction method, represented by a DNN, is used for prediction, along with commonly used logistic regression (LR), anecdotal neural networks (ANN), and XGBoost (XGB).
[0051] The experimental results are shown in Table 5. The method based on LLM and BERT achieved the best results on all datasets and classification methods, except for DNN on the T-2 sample set and XGB on the T-1 sample set. This indicates that the triple information extracted by the large language model does indeed contain more valuable information and reduces the interference of other redundant information on the prediction model. Moreover, the large language model is pre-trained on a large amount of historical data, which enhances its ability to understand context and has a significant advantage in capturing deep semantic information of text. Although the value of features decays as the latency of feature data increases, MD&A still contains important information such as the company's future development direction, which will be reflected in the company's future stock price and financial data.
[0052] SFL's performance is similar to the LLM-BERT method in many cases. Although it doesn't extract deep semantic information, it has a strong ability to mine the grammar and intonation of text, outperforming other non-LLM methods in most situations. Similar to existing research findings, it shows that the more standardized the grammar and the clearer the expression in MD&A, the clearer the signal it conveys to the outside world, and the better it can eliminate information asymmetry between external investors and the company. In addition, BERT outperforms LDA and SENTIMENT in most cases, further demonstrating the advantages of deep models in text semantic understanding. LLM-BERT's performance is significantly better than BERT, indicating that semantic extraction of the entire text will bring more redundant features and may even impair the predictive ability of classification models. Finally, LDA consistently performs the worst, indicating that simply extracting the topic information of the text cannot adequately understand the contextual information. Therefore, in financial risk management, managers should not only focus on external information but also combine internal information for targeted information disclosure. Furthermore, in the process of information disclosure, attention should be paid not only to grammar and intonation but also to the transmission of key information, avoiding the release of ineffective or ambiguous information that could introduce new risk factors.
[0053] Furthermore, triplet information extracted through language models (LLM) is superior to features extracted directly from text. Triple-based feature extraction methods can effectively capture key features that improve model performance. From a linguistic and syntactic perspective, triplet information extraction can more clearly reflect managers' views on the company's current operational status, risk factors, and future development trends.
[0054] Table 5 Comparison of the effects of various text feature extraction methods To verify the superiority of the MIFN model proposed in this invention in ST probability calculation and prediction, this invention further compares the prediction performance of the MIFN model with common LR, ANN, and XGB models. The comparison results are shown in Table 6. As can be seen from Table 6, the MIFN model proposed in this invention outperforms all other models in terms of AUC, KS, and F2 scores, which proves the effectiveness and robustness of the proposed model.
[0055] Table 6: Comparison of the performance of various prediction methods On the other hand, embodiments of the present invention also provide an enterprise financial risk prediction system based on multi-source feature fusion. This system uses the above-mentioned enterprise financial risk prediction method based on multi-source feature fusion, including: The data acquisition module is used to acquire social media comment data, MD&A data, and financial data. The data analysis module is used to perform semantic sentiment analysis on social networking site comment data using the FinBERT algorithm to obtain a first statistical feature set, to perform semantic sentiment analysis on MD&A data using the LLM-BERT model to obtain a second statistical feature set and a semantic feature set, and to calculate financial ratio data based on financial data. A model building module, which is used to build an MFIFN model; The prediction module is used to calculate the ST probability of an enterprise using the MFIFN model based on a first statistical feature set, a second statistical feature set, a semantic feature set, and financial ratio data, and to judge the financial risk of the enterprise based on the ST probability.
[0056] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A method for predicting corporate financial risk based on multi-source feature fusion, characterized in that: include: We acquire social media comment data and use the FinBERT algorithm to perform semantic sentiment analysis on the social media comment data to obtain the first statistical feature set. Obtain MD&A data, use the LLM-BERT model to perform semantic sentiment analysis on the MD&A data, and obtain the second statistical feature set and semantic feature set; Obtain financial data and calculate financial ratios based on the financial data; A MFIFN model is constructed, which is based on the first statistical feature set, the second statistical feature set, the semantic feature set, and financial ratio data. The MFIFN model is used to calculate the ST probability of a company, and the financial risk of the company is judged based on the ST probability.
2. The enterprise financial risk prediction method based on multi-source feature fusion as described in claim 1, characterized in that: Methods for using the FinBERT algorithm to perform semantic sentiment analysis on social networking site comment data and obtain the first statistical feature set include: The FinBERT algorithm was used to classify the sentiment of social media comment data, resulting in positive sentiment group, neutral sentiment group and negative sentiment group; Based on the data volume of the positive sentiment group and the negative sentiment group, calculate the first relative positivity and the first overall positivity between the positive sentiment group and the negative sentiment group; Calculate the percentage of data in the positive sentiment group, neutral sentiment group, and negative sentiment group relative to the total data volume; The data volume and corresponding data proportion of the positive sentiment group, neutral sentiment group, and negative sentiment group, the total data volume of social networking site comments, the first relative positivity, and the first overall positivity are combined into the first statistical feature set.
3. The enterprise financial risk prediction method based on multi-source feature fusion as described in claim 1, characterized in that: Methods for using the LLM-BERT model to perform semantic sentiment analysis on MD&A data and obtain the second statistical feature set and semantic feature set include: The LLM model was used to extract sentiment features from the MD&A data, resulting in positive, neutral, and negative triples. The BERT algorithm was used to perform semantic analysis on positive triples, neutral triples, and negative triples respectively, resulting in positive semantic feature sets, neutral semantic feature sets, and negative semantic feature sets. We obtain the semantic feature set by weighting all features in the positive semantic feature set, the neutral semantic feature set, and the negative semantic feature set. Based on the amount of data in the positive and negative triplets, calculate the second relative positivity and the second overall positivity between the positive triplets and the triadic sentiment groups; Calculate the percentage of data in positive triples, neutral triples, and negative triples relative to the total data volume; The data volume and corresponding data proportion in the positive, neutral, and negative triplet groups, the total data volume of the MD&A data, the second relative positivity, and the second overall positivity are combined to form the second statistical feature set.
4. The enterprise financial risk prediction method based on multi-source feature fusion as described in claim 2, characterized in that: The formula for calculating the first relative positivity is as follows: in, and These represent the data volume in the positive sentiment group and the negative sentiment group within the time period t.
5. The enterprise financial risk prediction method based on multi-source feature fusion as described in claim 1, characterized in that: Methods for calculating the ST probability of a firm using the MFIFN model include: Feature extraction is performed on the first statistical feature set, the second statistical feature set, the semantic feature set, and the financial ratio data respectively. All extracted features are then concatenated and fused to obtain the fused features. Based on the fusion features, the Focal loss algorithm is used to calculate the ST probability of the enterprise.
6. The enterprise financial risk prediction method based on multi-source feature fusion as described in claim 5, characterized in that: The methods for extracting features from the first statistical feature set, the second statistical feature set, the semantic feature set, and the financial ratio data, and then concatenating and fusing all extracted features to obtain the fused features include: Fully connected layers are used to extract features from the first statistical feature set, the second statistical feature set, and the financial ratio, respectively, to obtain the corresponding first shallow feature, second shallow feature, and third shallow feature. The LSTM model is used to extract features from the semantic feature set to obtain temporal features; A dense layer is used to extract features from the first shallow features, the second shallow features, the third shallow features, and the temporal features to obtain the corresponding deep features. Then, all deep features are concatenated and a linear attention mechanism is used to extract features to obtain the first attention feature. The second attention feature is obtained by using a linear attention mechanism to extract features from the temporal features. The first attention feature and the second attention feature are concatenated to obtain the fused feature.
7. The enterprise financial risk prediction method based on multi-source feature fusion as described in claim 6, characterized in that: When using the LSTM model to extract features from a semantic feature set, multiple time-series features are obtained at different time steps. The time-series features extracted at the last time step are then imported into a dense layer for feature extraction to obtain the corresponding deep features.
8. The enterprise financial risk prediction method based on multi-source feature fusion as described in claim 7, characterized in that: The temporal features extracted at each time step are concatenated, and then a linear attention mechanism is used to extract features to obtain the second attention features.
9. A corporate financial risk prediction system based on multi-source feature fusion, characterized in that: The system uses a multi-source feature fusion-based enterprise financial risk prediction method as described in any one of claims 1-8, comprising: The data acquisition module is used to acquire social media comment data, MD&A data, and financial data. The data analysis module is used to perform semantic sentiment analysis on social networking site comment data using the FinBERT algorithm to obtain a first statistical feature set, to perform semantic sentiment analysis on MD&A data using the LLM-BERT model to obtain a second statistical feature set and a semantic feature set, and to calculate financial ratio data based on financial data. A model building module, which is used to build an MFIFN model; The prediction module is used to calculate the ST probability of an enterprise using the MFIFN model based on a first statistical feature set, a second statistical feature set, a semantic feature set, and financial ratio data, and to judge the financial risk of the enterprise based on the ST probability.