Credit debt price fluctuation prediction method and device based on public opinion analysis

Abstract

The invention provides a credit debt price fluctuation prediction method and prediction device based on public opinion analysis, and the method comprises the steps: obtaining first public opinion data and second public opinion data, determining a first event tag and a second event tag based on the first public opinion data and the second public opinion data, and determining a risk conduction coefficient according to correlation; integrating the second public opinion data and the first public opinion data according to the risk conduction coefficient to obtain a data wide table; processing the data wide table through an APRIORI algorithm and an LSTM algorithm to obtain an event combination set and an event chain; and predicting a price transaction probability through a logistic regression model or predicting a first preset risk event occurrence probability through an MLP model based on the event combination set and the event chain, and generating target alarm information under the condition that any probability exceeds a limit. The method solves the problem that in the prior art, public opinion data mining is insufficient in a credit debt evaluation method, so that the accuracy of an evaluation result is low, and enterprise financial loss is caused.

Description

Technical Field

[0001] This invention relates to the field of bond risk early warning technology, and more specifically, to a method, device, computer-readable storage medium, and risk monitoring system for predicting credit bond price fluctuations based on public opinion analysis. Background Technology

[0002] Credit bonds refer to bonds issued by enterprises, with principal and interest repayment cash flows determined based on the enterprise's credit agreement. Therefore, they are significantly affected by the enterprise's credit. However, the credit risk of enterprises is often hidden, making it difficult to fully identify the credit risk of the bond issuer.

[0003] Therefore, the external credit rating of bond issuers is a crucial reference for risk management. Traditional methods rely on corporate financial statements to build models, analyze the credit risk of bond issuers, and ultimately realize cash flow. However, financial statements reflect a company's financial status and operating results over a specific period, which carries the risk of untimely and distorted information. Furthermore, market environments and industry trends change rapidly, making risk monitoring based on financial statements inherently lagging, and their accuracy depends on the honesty of the financial statements. Similarly, modeling based on valuation and market transaction data is also a form of ex-post monitoring, and its accuracy depends on the timeliness and completeness of the data, as well as the authority of the valuation agency, thus suffering from the same drawbacks.

[0004] To address the aforementioned shortcomings, existing technologies propose using the proportion of negative words in news data (public opinion events) of the bond issuer as the primary basis for constructing a multi-dimensional risk characteristic system based on keywords. This represents a significant improvement over traditional methods. However, research has revealed the following drawbacks in the utilization of public opinion data in existing technologies:

[0005] 1) A risk feature system is constructed based on negative words as keywords. It can be relatively accurate in analyzing serious public opinion events. However, due to the difficulty in constructing the system itself and the data limitations on the basis for construction, it focuses on analyzing public opinion events that have already had obvious impact. It lacks effective means of analyzing the possible impact of minor public opinion events and potential undeveloped public opinion events.

[0006] 2) The analysis only focuses on the news data of the bond issuer, and the analysis object is limited to the bond issuer itself. The analysis of a single entity lacks the analysis of the derivative impact of public opinion and the interference of the chain. For example, the impact of public opinion events of the company's upstream and downstream companies on the company's credit.

[0007] 3) The prediction is limited to credit bond ratings, which reflect the probability of credit bond defaults through credit bond rating adjustments, but ignores the losses caused by changes in credit bond prices themselves.

[0008] In summary, existing technologies struggle to provide accurate and effective methods for predicting the risks of credit bonds, resulting in poor performance in mitigating losses. Summary of the Invention

[0009] The main purpose of this application is to provide a method, device, computer-readable storage medium, and risk monitoring system for predicting credit bond price fluctuations based on public opinion analysis, so as to at least solve the problem that existing credit bond assessment methods are insufficient in mining public opinion data, resulting in low accuracy of assessment results and causing financial losses to enterprises.

[0010] To achieve the above objectives, according to one aspect of this application, a method for predicting credit bond price fluctuations based on public opinion analysis is provided, comprising: acquiring news data of the bond issuer to obtain first public opinion data; acquiring news data of related entities of the bond issuer to obtain second public opinion data; determining public opinion events of the bond issuer based on the first public opinion data to obtain a first event tag; determining the public opinion events of the related entities based on the second public opinion data to obtain a second event tag; determining a risk transmission coefficient based on the correlation between the bond issuer and the related entities, wherein the risk transmission coefficient is used to characterize the degree of influence between the bond issuer and the related entities; associating the second event tag and the second public opinion data with the first event tag and the first public opinion data based on the risk transmission coefficient to obtain a data wide table; and processing the data wide table using the APRIORI algorithm. An event set is obtained, which includes at least a combination of public opinion events with risk orientation. The data wide table is processed using an LSTM algorithm to obtain an event chain, which is used to characterize the changing trend of the public opinion event within a predetermined first time period after the current moment. Based on the event set and the event chain, a logistic regression model is used to predict the probability of price fluctuation of the bond issuer, or an MLP model is used to predict the probability of a first predetermined risk event. If the probability of price fluctuation or the probability of risk is greater than the corresponding predetermined threshold, a target alarm message is generated. The logistic regression model is trained based on the combination of public opinion events, the changing trend of the public opinion events, and the fluctuation of credit bond prices in historical periods. The MLP model is trained based on whether the event chain and the first predetermined risk event have occurred in historical periods.

[0011] Optionally, the second event tag and the second public opinion data are associated with the first event tag and the first public opinion data according to the risk transmission coefficient to obtain a wide data table, including: performing word segmentation on the first public opinion data using NLP technology, and filtering out conjunctions, stop words, and part-of-speech tags in the segmented first public opinion data to obtain third public opinion data; performing word segmentation on the second public opinion data using the NLP technology, and filtering out conjunctions, stop words, and part-of-speech tags in the segmented second public opinion data to obtain fourth public opinion data; converting the third public opinion data into a feature vector using the TF-IDF algorithm to obtain a first target vector; and converting the fourth public opinion data into a feature vector using the TF-IDF algorithm. The data is converted into feature vectors to obtain a second target vector; the first target vector is clustered to obtain multiple first target clusters, each corresponding to a first event label; the second target vector is clustered to obtain multiple second target clusters, each corresponding to a second event label; based on the first event label, the second event label, and the risk transmission coefficient, the second target clusters are aggregated with the corresponding first target clusters to obtain multiple third target clusters; the corresponding event labels are determined according to each third target cluster to obtain target event labels; the target event labels are used as the headers to fill in the corresponding third target clusters to obtain the wide data table.

[0012] Optionally, the data wide table is processed using the APRIORI algorithm to obtain an event combination set, including: determining the risk level of the corresponding public opinion event based on the third target cluster corresponding to each target event label, and obtaining the target risk level; filtering out the third target clusters in the data wide table whose target risk level is greater than a preset level, and obtaining an updated data wide table; and processing the updated data wide table using the APRIORI algorithm to obtain the event combination set.

[0013] Optionally, the event combination set is obtained by processing the updated wide data table using the APRIORI algorithm, including: filtering public opinion event combinations with support greater than a preset support at different scales based on the updated wide data table to obtain a first frequent itemset; constructing corresponding association rules for each public opinion event combination in the first frequent itemset to obtain multiple association rule sets; filtering the first frequent itemset based on the association rule sets to obtain the event combination set, wherein the association rules in the association rule sets are used to characterize the causal relationship between the combination of public opinion events and the first preset risk event.

[0014] Optionally, the first frequent itemset is obtained by filtering combinations of public opinion events with support greater than a preset support at different scales according to the updated wide data table, including: using a single public opinion event as a basic element, converting all public opinion events in the updated wide data table into the basic elements in the dataset to obtain candidate frequent itemsets; calculating the support of each basic element in the candidate frequent itemset according to the updated wide data table to obtain a first support; constructing a second frequent itemset according to the basic elements corresponding to the first support being greater than the preset support; and a first update step, whereby the basic elements in the second frequent itemset are randomly matched pairwise. The first update step involves using the combination of elements as the basic elements to update the candidate frequent itemset, resulting in an updated candidate frequent itemset where any two basic elements in the updated candidate frequent itemset are inconsistent. The second update step involves recalculating the support based on the updated candidate frequent itemset to obtain a second support, and updating the second frequent itemset according to the second support. The first and second update steps are repeated sequentially until the second support corresponding to all basic elements in the updated candidate frequent itemset is less than the preset support. The last updated second frequent itemset is then determined as the first frequent itemset.

[0015] Optionally, based on each of the public opinion event combinations in the first frequent itemset, corresponding association rules are constructed to obtain multiple association rule sets. The first frequent itemset is then filtered according to the association rule sets to obtain the event combination set. This includes: establishing association rules between each basic element in the first frequent itemset and each of the first preset risk events, wherein the association rules are based on the basic elements and conclude with the first preset risk events; calculating the corresponding confidence level for each association rule to obtain a target confidence level; calculating the lift of the association rules based on the target confidence level when the target confidence level is greater than a preset confidence level to obtain a target lift; constructing the association rule set based on the association rules whose target lift is greater than a preset lift; updating the first frequent itemset with the basic elements included in the association rule set, and merging the updated first frequent itemset and the association rule set to obtain the event combination set.

[0016] Optionally, news data of the bond issuer is obtained to obtain first public opinion data. Based on the first public opinion data, public opinion events of the bond issuer are determined to obtain first event tags. This includes: crawling news data of the bond issuer from multiple preset paths to obtain fifth public opinion data; cleaning the fifth public opinion data to obtain the first public opinion data; determining the category to which the first public opinion data belongs based on the first preset risk events existing in the business scenarios involved in the first public opinion data to obtain a first-level tag; and determining the category of the public opinion events to which the first public opinion data belongs that led to the first preset risk events to obtain a second-level tag. The first-level tags include financial risk, operational risk, management risk, and legal risk. The second-level tags include payment delays, product quality inspection issues, management personnel changes, litigation events, and supply chain changes. The first-level tags and the second-level tags are determined as the first event tags of the first public opinion data.

[0017] According to another aspect of this application, a credit bond price fluctuation prediction device based on public opinion analysis is provided. The device includes: a first acquisition unit, configured to acquire news data of the bond issuer to obtain first public opinion data; acquire news data of related entities to obtain second public opinion data; determine public opinion events of the bond issuer based on the first public opinion data to obtain a first event tag; determine the public opinion events of the related entities based on the second public opinion data to obtain a second event tag; and determine a risk transmission coefficient based on the association between the bond issuer and the related entities, the risk transmission coefficient being used to characterize the degree of influence between the bond issuer and the related entities; and a first processing unit, configured to associate the second event tag and the second public opinion data with the first event tag and the first public opinion data based on the risk transmission coefficient. The system employs a first processing unit to process data, resulting in a wide data table. A second processing unit processes the wide data table using the APRIORI algorithm to obtain an event combination set, which includes at least risk-oriented public opinion event combinations. A third processing unit processes the wide data table using the LSTM algorithm to obtain an event chain, which characterizes the changing trend of the public opinion event within a predetermined first time period after the current moment. A prediction unit uses a logistic regression model based on the event combination set and the event chain to predict the probability of price anomalies in the bond issuer. If the probability of price anomalies exceeds a preset threshold, a target alarm is generated. The logistic regression model is trained based on the public opinion event combinations, the changing trends of the public opinion events, and credit bond price fluctuations over historical periods.

[0018] According to another aspect of this application, a computer-readable storage medium is provided, the computer-readable storage medium including a stored program, wherein, when the program is executed, it controls any of the methods described in the device on which the computer-readable storage medium is located.

[0019] According to another aspect of this application, a risk monitoring system is provided, comprising: one or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs including methods for performing any one of the methods described.

[0020] Applying the technical solution of this application, in the aforementioned credit bond price fluctuation prediction method based on public opinion analysis, firstly, news data of the bond issuer is acquired to obtain first public opinion data; news data of related entities of the bond issuer is acquired to obtain second public opinion data; based on the first public opinion data, public opinion events of the bond issuer are determined to obtain first event tags; based on the second public opinion data, public opinion events of the related entities are determined to obtain second event tags; a risk transmission coefficient is determined based on the correlation between the bond issuer and the related entities, and the risk transmission coefficient is used to characterize the degree of influence between the bond issuer and the related entities; then, based on the risk transmission coefficient, the second event tags and the second public opinion data are associated with the first event tags and the first public opinion data to obtain a wide data table; subsequently, the wide data table is processed using the APRIORI algorithm to obtain event groups. The data set includes at least a combination of risk-oriented public opinion events. Then, the wide data table is processed using an LSTM algorithm to obtain event chains, which characterize the changing trends of the aforementioned public opinion events within a predetermined first time period after the current moment. Finally, based on the event set and the event chains, a logistic regression model is used to predict the probability of price anomalies in the bond issuer, or an MLP model is used to predict the probability of a first predetermined risk event occurring. If the probability of price anomalies or the probability of risk exceeds a corresponding predetermined threshold, a target warning is generated. The logistic regression model is trained based on the aforementioned public opinion event combinations, the changing trends of the aforementioned public opinion events, and credit bond price fluctuations within a historical period. The MLP model is trained based on whether the aforementioned event chains and the first predetermined risk event have occurred within a historical period. This application integrates public opinion data by mining the upstream and downstream relationships between bond issuers and related entities. Based on this integrated data, it uses APRIORI to mine the correlations between public opinion data, and employs the LSTM algorithm to predict the development trend of public opinion events over time to uncover potentially risky public opinion data. A logistic regression model is used to synthesize the correlations between public opinion data and the current trend of public opinion data to predict future credit bond price fluctuations that may be caused by changes in public opinion, thus avoiding losses. Compared to existing technologies that rely on the proportion of key public opinion and negative public opinion, this application improves upon existing technologies by expanding the dimensions of data analysis, introducing the mutual influence between public opinion data of different entities, mining the correlations between public opinion data, introducing the mutual influence between different public opinion events, and mining the development chain of public opinion events to avoid missing potential public opinion. Furthermore, by using a logistic regression model to synthesize the current public opinion situation resulting from the influence of different events and predicting the development trend based on the current public opinion events, this application solves the problem that existing credit bond assessment methods lack sufficient public opinion data mining, leading to low accuracy in assessment results and causing financial losses for enterprises. Attached Figure Description

[0021] Figure 1 A flowchart illustrating a method for predicting credit bond price fluctuations based on public opinion analysis, according to an embodiment of this application, is shown.

[0022] Figure 2 A flowchart illustrating a specific method for predicting credit bond price fluctuations based on public opinion analysis, according to another embodiment of this application, is shown.

[0023] Figure 3 A structural block diagram of a credit bond price fluctuation prediction device based on public opinion analysis, provided according to an embodiment of this application, is shown. Detailed Implementation

[0024] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0025] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0026] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this application described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0027] For ease of description, the following explains some of the nouns or terms used in the embodiments of this application:

[0028] TF-IDF (Term Frequency-Inverse Document Frequency) is a statistical method used to evaluate the importance of a word in a document set or corpus. This method has wide applications in information retrieval, document classification, keyword extraction, and other scenarios. Although TF-IDF itself is not directly used for clustering, it can serve as an important feature extraction step in clustering algorithms, especially in content-based text clustering.

[0029] Term frequency (TF): refers to the frequency of a word in a document. It can be expressed as the number of times a word appears in the document, or as relative term frequency (i.e., the number of times a word appears in the document divided by the total number of words in the document).

[0030] Inverse Document Frequency (IDF): Introduced to measure the general importance of a word. It is calculated by taking the inverse of the number of documents in which the word appears in the entire document set, and usually smoothed by taking the logarithm. If a word appears in most documents, its IDF value will be relatively low; conversely, if a word appears in only a few documents, its IDF value will be relatively high.

[0031] K-means: A widely used clustering algorithm, belonging to the category of unsupervised learning algorithms. Its goal is to divide n observations (or data points) into k clusters, such that the data points within each cluster have high similarity, while the data points between different clusters have low similarity.

[0032] LSTM (Long Short-Term Memory) is a special recurrent neural network (RNN) architecture used to address the vanishing or exploding gradient problems that traditional RNNs encounter when dealing with long-term dependencies. LSTM uses three "gates" (forget gate, input gate, and output gate) to control the flow of information, thereby enabling it to capture long-term dependencies in sequential data.

[0033] MLP (Multi-Layer Perceptron) is a feedforward artificial neural network model consisting of an input layer, one or more hidden layers, and an output layer. MLP is one of the simplest neural networks, yet it has been able to solve many complex classification and regression problems.

[0034] As described in the background section, existing technologies struggle to accurately and effectively predict the risks of credit bonds, resulting in poor loss mitigation. To address the issue of insufficient public opinion data mining in existing credit bond assessment methods, leading to low accuracy in assessment results and causing financial losses for enterprises, embodiments of this application provide a credit bond price fluctuation prediction method, prediction device, computer-readable storage medium, and risk monitoring system based on public opinion analysis.

[0035] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0036] This embodiment provides a method for predicting credit bond price fluctuations based on public opinion analysis, which runs on a mobile terminal, computer terminal, or similar computing device. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Also, although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order than that shown here.

[0037] Figure 1 This is a flowchart of a credit bond price fluctuation prediction method based on public opinion analysis according to an embodiment of this application.

[0038] like Figure 1 As shown, the method includes the following steps:

[0039] Step S201: Obtain news data of the bond issuer to obtain first public opinion data; obtain news data of related entities of the bond issuer to obtain second public opinion data; determine the public opinion events of the bond issuer based on the first public opinion data to obtain first event tags; determine the public opinion events of the related entities based on the second public opinion data to obtain second event tags; determine the risk transmission coefficient based on the relationship between the bond issuer and the related entities, and the risk transmission coefficient is used to characterize the degree of influence between the bond issuer and the related entities.

[0040] Specifically, such as Figure 2 As shown, in step 1, news data of the bond issuer and related entities (enterprises, individuals or institutions that have dealings with the bond issuer) are collected to obtain the first public opinion data and the second public opinion data mentioned above. Further, the event tags of the public opinion data are determined based on the risk event type to which the public opinion data belongs.

[0041] In practice, the risk transmission coefficient is determined based on the relationship between the aforementioned bond issuers and related entities, including:

[0042] The aforementioned bond issuers and related entities are represented by nodes. The nodes corresponding to the directly related bond issuers and related entities are connected by solid lines, and the nodes corresponding to the indirectly related bond issuers and related entities are connected by dashed lines to obtain the relationship graph.

[0043] Based on the above correlation graph, a risk transmission model is constructed based on a Bayesian neural network. The risk transmission model is trained based on the news data and credit risk of the above-mentioned bond issuers and related entities before the current time until the risk transmission model converges. The weights of the connecting lines between the above-mentioned nodes in the converged risk transmission model are determined as the risk transmission coefficients. The risk transmission coefficients corresponding to the solid lines are greater than the risk transmission coefficients corresponding to the dashed lines.

[0044] Understandably, the impact between the bond issuer and related entities is not uniformly distributed, and the introduction of the risk transmission coefficient aims to quantify the degree of this impact. By constructing a correlation graph and training a risk transmission model, the likelihood of risk transmission between different entities can be accurately assessed, providing more precise data support for subsequent risk prediction and resolving the ambiguity in risk transmission analysis in traditional methods.

[0045] Through the above embodiments, by constructing a correlation graph and a risk transmission model, a quantitative analysis of the risk transmission path between the bond issuer and related entities is achieved, solving the problem of ambiguity and difficulty in quantifying risk transmission relationships in traditional risk assessment. The application of Bayesian neural networks enables the model to automatically adjust the connection weights between nodes, i.e., the risk transmission coefficients, based on historical data, thereby more accurately reflecting the intensity of risk transmission between different entities. The distinction between solid and dashed lines further refines the identification of risk transmission paths, helping financial institutions to more accurately assess the sources of risk in credit bond price fluctuations. This not only improves the accuracy and timeliness of risk warnings in the credit bond market but also reveals the complex relationships between events, providing financial institutions with in-depth risk insights.

[0046] Step S202: Based on the risk transmission coefficient, associate the second event label and the second public opinion data with the first event label and the first public opinion data to obtain a wide data table;

[0047] Specifically, such as Figure 2 As shown, step 2 is news event clustering, which involves integrating the first and second public opinion data mentioned above, and using clustering to integrate public opinion data corresponding to the same event into the same label.

[0048] Step S203: Process the above wide data table using the APRIORI algorithm to obtain an event combination set, which includes at least a combination of public opinion events with risk orientation.

[0049] Specifically, such as Figure 2 As shown, step 3 is public opinion trend analysis, which uses the APRIORI model to quantify the trend of individual risk events that have low risk but, when combined, have a high probability of causing serious defaults or price changes, in order to uncover the public opinion trend of minor events.

[0050] Step S204: The above data wide table is processed by the LSTM algorithm to obtain the event chain. The event chain is used to characterize the change trend of the above public opinion event within a preset first time period after the current moment.

[0051] Specifically, such as Figure 2 As shown, step 4 is event chain prediction, which is to predict the event chain based on the time series using the LSTM model. It can be understood that the model predicts the future public opinion events at a preset step size based on the current public opinion events, and forms an event chain through continuous iteration.

[0052] In practical implementation, the above wide data table is processed using the LSTM algorithm to obtain the event chain, including:

[0053] Obtain the aforementioned public opinion events with any first preset duration before the current time and convert them into vector form to obtain the third target vector;

[0054] The third target vector is divided into a training set and a validation set according to a preset ratio. An LSTM model is trained based on the training set and the performance of the LSTM model is verified using the validation set.

[0055] If the verification is successful, the above wide data table is used as the initial input data. The above LSTM model is used to predict the public opinion data after a second preset time interval. The prediction results are repeatedly input into the above LSTM model to predict the next prediction result until the last prediction result is separated from the above initial input data by the above first preset time interval. All the output results of the above LSTM model are connected in chronological order to obtain the above event chain.

[0056] Understandably, event chains can reflect the dynamic relationships of public opinion events over time, which is crucial for predicting future event trends. Constructing event chains using the LSTM algorithm can capture the temporal dependencies between events, providing a dynamic perspective for risk prediction and addressing the problem that static analysis cannot predict the direction of risk evolution.

[0057] Through the above embodiments, by constructing event chains, the future trends of public opinion events can be predicted, solving the problems of weak event sequence analysis capabilities and low prediction accuracy in traditional prediction models. The application of the LSTM model enables the model to capture the temporal dependencies between events, improving prediction accuracy. By dividing the training and validation sets, this approach ensures the model's generalization ability and prediction performance, avoiding overfitting. Furthermore, the construction of event chains not only predicts the future development trends of public opinion events but also reveals the potential connections between events from a temporal perspective.

[0058] Step S205: Based on the above event combination set and the above event chain, predict the probability of price fluctuation of the above bond issuer through a logistic regression model or predict the probability of the occurrence of the first preset risk event through an MLP model based on the above event chain. If the above price fluctuation probability or the above risk probability is greater than the corresponding preset threshold, generate target alarm information. The above logistic regression model is trained based on the above public opinion event combination, the changing trend of the above public opinion events and the fluctuation of credit bond prices in the historical period. The above MLP model is trained based on whether the above event chain and the above first preset risk event have occurred in the historical period.

[0059] Specifically, such as Figure 2 As shown, step 5, anomaly prediction, includes the prediction of price anomalies and the prediction of risk events. Price anomaly prediction is achieved by combining the public opinion analysis results obtained in step 3 and the event chain obtained in step 4 with a logistic regression model, expanding it into the final modeling dataset according to feature derivation rules, and then predicting credit bond price anomalies and issuing warnings based on the prediction results. Risk event prediction is achieved by using an MLP model based on the public opinion analysis results integrated in step 3, determining the probability that small events combined will lead to risk events in the future, and issuing warnings based on the prediction results.

[0060] In practical implementation, based on the above event chain, the probability of the occurrence of the first preset risk event is predicted using an MLP model. If the above risk probability is greater than the corresponding preset threshold, target alarm information is generated, including:

[0061] The above event chain is input into the above MLP model to predict the probability of each of the above first preset risk events occurring after the above first preset time period, thereby obtaining multiple first target probabilities;

[0062] Based on the comparison between the first target probability and the first preset probability corresponding to each of the first preset risk events, if the first target probability is greater than the corresponding first preset probability, the first preset risk event is marked, and the first preset probability corresponds one-to-one with the first preset risk event.

[0063] The aforementioned target alarm information is generated based on the first preset risk event that has been marked.

[0064] Understandably, the MLP model's prediction of the probability of risk events is a further extension based on event chain analysis. By inputting the event chain into the MLP model, the probability of a specific risk event can be predicted while considering the dynamic changes of events, thus solving the problem of traditional prediction methods lacking foresight regarding the future development of events.

[0065] Through the above embodiments, the probability of risk events is predicted using an MLP model, and an alarm message is generated when the probability exceeds a preset threshold. This method combines the prediction results of the event chain, further improving the timeliness and accuracy of risk warnings. Specifically, by processing the event chain through the MLP model, an accurate prediction of the probability of the first preset risk event is achieved, solving the problems of inaccurate probability quantification and imperfect early warning mechanisms in traditional risk prediction. The calculation of the first target probability enables the model to quantify the impact of the event chain on the risk of credit bond price volatility, providing financial institutions with data-supported risk warnings.

[0066] Furthermore, before inputting the aforementioned event chain into the aforementioned MLP model, the method may also include:

[0067] Based on historical public opinion data, the above LSTM model is used to predict and obtain the public opinion data sequence.

[0068] The aforementioned public opinion data sequence and the aforementioned historical public opinion data are determined as the input data of the aforementioned MLP model. Label data is determined based on the aforementioned input data. The aforementioned label data is used to characterize whether the aforementioned first preset risk event corresponding to the aforementioned input data has occurred.

[0069] The input data and the label data are divided into a training set and a test set according to a preset ratio. The MLP model is trained based on the training set and the trained MLP model is tested on the test set.

[0070] If the test passes, the MLP model is considered to have completed training. If the test fails, at least one of the following parameters of the MLP model is adjusted: learning rate, regularization parameter, number of hidden layers, and number of neurons, and the model is retrained until the MLP model is trained. During training, DROPOUT technique and L1 or L2 regularization are used to prevent the MLP model from overfitting.

[0071] Understandably, training an MLP model requires a large amount of high-quality historical public opinion data. To avoid overfitting, adjusting the learning rate, regularization parameters, and using the dropout technique can improve the model's generalization ability, ensuring that the model not only performs well on the training data but also makes accurate predictions on new, unseen public opinion data, thus solving the problem of weak model generalization.

[0072] Through the above embodiments, the accuracy of the model is achieved while ensuring its stability and reliability.

[0073] In practical implementation, based on the aforementioned event set and event chain, a logistic regression model is used to predict the probability of price anomalies in the aforementioned bond issuers, including:

[0074] The above event set and event chain are concatenated into the same matrix to obtain the input feature matrix;

[0075] The above-mentioned price fluctuation probability is obtained by processing the above-mentioned input feature matrix through the trained logistic regression model.

[0076] Understandably, logistic regression models can handle binary classification problems and are suitable for predicting the likelihood of abnormal fluctuations in credit bond prices. By concatenating event combinations and event chains into an input feature matrix, logistic regression models can integrate multi-source risk information, providing a more comprehensive risk assessment and solving the problem of insufficient prediction by a single risk indicator.

[0077] Through the above embodiments, by concatenating event sets and event chains into an input feature matrix, and processing it using a trained model, a prediction result is obtained. This method can comprehensively consider multiple risk factors and provide comprehensive risk warnings.

[0078] Furthermore, before processing the input feature matrix using the trained logistic regression model to obtain the price anomaly probability, the method may further include:

[0079] The above-mentioned public opinion events and their changing trends within a historical period are transformed into vector form and then concatenated to obtain the target matrix. The fluctuations in credit bond prices are then transformed into vector form to obtain the benchmark output result.

[0080] Based on the above target matrix and the above benchmark output results, the dataset is divided into a training set and a test set according to a preset ratio. The above logistic regression model is trained based on the above training set, and the trained logistic regression model is evaluated on the above test set by one or more of the following: accuracy, precision, recall, F1 score, and AUC-+ROC curve, to obtain a performance evaluation score.

[0081] If the performance evaluation score is greater than the preset threshold, the logistic regression model is determined to be trained successfully. If the performance evaluation score is less than the preset threshold, the weights of the combination of public opinion events and the changing trends of the public opinion events in the splicing process are adjusted, and the training is repeated until the logistic regression model is trained successfully.

[0082] Understandably, training and performance evaluation of logistic regression models are crucial steps in ensuring their predictive ability. By setting performance evaluation thresholds, truly effective models can be selected, avoiding prediction bias caused by poor model performance. This mechanism guarantees the model's accuracy and robustness, solving the problem of model performance instability.

[0083] Through the above embodiments, by adjusting model parameters and feature weights, the model performance is optimized, ensuring the accuracy and reliability of model predictions.

[0084] This embodiment integrates public opinion data by mining the upstream and downstream relationships between the bond issuer and related entities. Based on this integrated data, the APRIORI algorithm is used to mine the correlations between public opinion data, and the LSTM algorithm predicts the development trend of public opinion events over time to uncover potentially risky public opinion data. A logistic regression model is used to synthesize the correlations between public opinion data and the current trend of public opinion data to predict future credit bond price fluctuations that may be caused by changes in public opinion, thus avoiding losses. Compared to existing technologies that rely on the proportion of key public opinion and negative public opinion, this application improves upon existing technologies by expanding the dimensions of data analysis, introducing the mutual influence between public opinion data of different entities, mining the correlations between public opinion data, introducing the mutual influence between different public opinion events, and mining the development chain of public opinion events to avoid missing potential public opinion. Furthermore, by using a logistic regression model to synthesize the current public opinion situation resulting from the influence of different events and predicting the development trend based on the current public opinion events, this invention solves the problem that existing credit bond assessment methods lack sufficient public opinion data mining, leading to low accuracy in assessment results and causing financial losses for enterprises.

[0085] In order to integrate the aforementioned first and second public opinion data, in an optional implementation, step S202 includes:

[0086] Step S2021: The first public opinion data is segmented using NLP technology, and the connecting words, stop words and part-of-speech tags in the segmented first public opinion data are removed to obtain the third public opinion data. The second public opinion data is segmented using NLP technology, and the connecting words, stop words and part-of-speech tags in the segmented second public opinion data are removed to obtain the fourth public opinion data.

[0087] Specifically, NLP technology is used to segment the first and second public opinion data after data cleaning. Furthermore, the connecting words, stop words, and part-of-speech tags after segmentation are removed, which simplifies the computational complexity and improves the accuracy of subsequent analysis.

[0088] Step S2022: Convert the third public opinion data into a feature vector using the TF-IDF algorithm to obtain the first target vector; convert the fourth public opinion data into a feature vector using the TF-IDF algorithm to obtain the second target vector.

[0089] Specifically, the TF and IDF values ​​of each word in the corresponding document are calculated and multiplied to obtain the importance of each word in a certain document. Then, each document is represented as a feature vector, with each dimension of the vector corresponding to the TF-IDF value of a word.

[0090] Step S2023: Cluster the first target vector to obtain multiple first target clusters, and each first target cluster corresponds to a first event label.

[0091] Specifically, the number of clusters (k) is set by the first event label mentioned above, and then the feature vectors are clustered by the K-means clustering algorithm to integrate different texts of the same public opinion event and obtain the first target cluster.

[0092] Step S2024: Cluster the second target vector to obtain multiple second target clusters, and each second target cluster corresponds to a second event label.

[0093] Specifically, the same processing is performed on the second target vector to obtain the second target cluster.

[0094] Step S2025: Based on the first event label, the second event label, and the risk transmission coefficient, the second target cluster is aggregated with the corresponding first target cluster to obtain multiple third target clusters. The corresponding event label is determined according to each third target cluster to obtain the target event label.

[0095] Specifically, based on the correlation between the first event label and the second event label, the vectors in the clusters under the labels are integrated according to the risk transmission coefficient to obtain the third target cluster. The type of public opinion event to which the cluster belongs is determined based on the integrated third target cluster to obtain the target event label.

[0096] In specific implementation, based on the aforementioned first event label, the aforementioned second event label, and the aforementioned risk transmission coefficient, the aforementioned second target cluster is aggregated with the corresponding aforementioned first target cluster to obtain multiple third target clusters, including:

[0097] Based on the above correlation map, the correlation relationship between the first event label and the second event label and the risk transmission coefficient are determined;

[0098] Based on the aforementioned risk transmission coefficients, the feature vectors in the corresponding second target clusters are transformed, and the transformed feature vectors in the second target clusters are concatenated with the feature vectors in the first target clusters that have the aforementioned correlation to obtain the third target clusters.

[0099] Understandably, the relationships between public opinion events are complex and diverse, and analyzing a single event can easily overlook potential chain reactions. By constructing a relationship graph, the connections and impacts between events can be clearly shown. The aggregation mechanism based on the risk transmission coefficient can more accurately assess the overall risk of related events, solving the problem of the one-sidedness of analyzing isolated events.

[0100] Step S2026: Fill in the third target cluster with the target event label as the table header to obtain the wide data table.

[0101] Understandably, public opinion data is massive and highly unstructured. Directly using raw data leads to inefficient model training and prediction, and makes it difficult to capture truly important information. By using NLP techniques for word segmentation, stop word removal, and part-of-speech tagging, and then combining this with the TF-IDF algorithm to transform the text into feature vectors, key information can be effectively extracted and quantified, improving the processing power and prediction accuracy of subsequent models, while reducing data noise.

[0102] Through the above embodiments, by employing Natural Language Processing (NLP) technology to preprocess public opinion data, this solution can effectively extract keywords and phrases directly related to the credit risk of bond issuers, remove irrelevant words, and improve the efficiency and accuracy of subsequent analysis. The application of the TF-IDF algorithm further quantifies the importance of keywords in public opinion data, providing a data foundation for event clustering. Cluster analysis enables similar public opinion events to be categorized, forming event labels, reducing noise and computational complexity caused by repeated calculations of identical events. The use of risk transmission coefficients ensures that public opinion events related to relevant entities can be reasonably aggregated according to their potential impact on bond issuers. The constructed wide data table provides comprehensive event information for subsequent machine learning model training.

[0103] In order to analyze the public opinion situation, in one optional implementation, step S203 includes:

[0104] Step S2031: Determine the risk level of the aforementioned public opinion event based on the third target cluster corresponding to each of the aforementioned target event tags, and obtain the target risk level;

[0105] Specifically, by evaluating the public opinion events in the aforementioned third target cluster, the risk level of the corresponding public opinion events is determined, thereby eliminating the low-reference event label of "major public opinion events" (major public opinion events directly reflect the risks, so they are not included in the public opinion situation analysis and can be directly used for subsequent analysis).

[0106] In practical implementation, the risk level of the corresponding event is determined based on the third target cluster corresponding to each of the above-mentioned target event labels, thus obtaining the target risk level, including:

[0107] Based on the feature vectors in the third target cluster mentioned above, the target labels and sentiment intensity scores are obtained by labeling them using a sentiment analysis algorithm. The target labels include positive labels, neutral labels and negative labels.

[0108] Based on the TF-IDF algorithm, the feature vectors in the third target cluster are calculated to obtain the importance score. Feature vectors with an importance score greater than the preset value are selected to obtain the key vectors. The importance score represents the degree of importance of the feature vector in the above public opinion event.

[0109] Based on the aforementioned key vectors, sentiment intensity scores, and target labels, the corresponding preset weights are queried to obtain multiple target weights;

[0110] The aforementioned key vectors, sentiment intensity scores, target labels, and corresponding target weights are input into the risk quantification model to obtain the aforementioned target risk level. The aforementioned risk quantification model is trained based on the aforementioned key vectors, sentiment intensity scores, target labels, and preset risk levels extracted from news data within historical time periods.

[0111] Understandably, risk level assessment is a core component of risk early warning. By employing sentiment analysis, TF-IDF algorithms, and risk quantification models, the impact of an event can be comprehensively evaluated from different perspectives, providing a quantitative basis for risk early warning. This multi-dimensional assessment method improves the scientific rigor and accuracy of risk prediction, overcoming the limitations of single-indicator assessments.

[0112] Through the above embodiments, a risk quantification model is used to assess the risk level of events. By combining key vectors, sentiment intensity scores, and target labels, a method for quantifying the degree of risk is provided, which helps to more accurately identify and warn of potential risk events.

[0113] Step S2032: Filter out the third target clusters in the above data wide table that correspond to the target risk level being greater than the preset level, and obtain the updated data wide table;

[0114] Step S2033: Based on the updated wide data table, process it using the APRIORI algorithm to obtain the event combination set.

[0115] Specifically, based on a wide data table after removing major public opinion events, the APRIORI model is used to quantify the public opinion situation of individual events with low importance, but which, when combined, would likely lead to serious default risk events.

[0116] Understandably, high-risk events have a significant impact on corporate bond price fluctuations, but they are relatively few in number and easily detected by existing early warning mechanisms. In contrast, combinations of low-risk events can lead to unexpected cumulative risk effects. By using the APRIORI algorithm to screen for high-risk events and focusing on combinations of low-risk events, potential "black swan" events can be identified, improving the comprehensiveness and timeliness of risk warnings.

[0117] Through the above embodiments, the data wide table was further optimized by screening for risk levels, ensuring high-quality model training data and solving the problem of difficulty in filtering effective information from a large number of public opinion events. The application of the APRIORI algorithm can mine event combinations with high support from the updated data wide table.

[0118] To tap into the aggregation effect of subtle public opinion events, in one optional implementation, step S2033 includes:

[0119] Step S20331: Based on the above updated data wide table, filter the combinations of public opinion events with support greater than the preset support at different scales to obtain the first frequent itemset;

[0120] Step S20332: Based on each of the above-mentioned public opinion event combinations in the first frequent item set, construct corresponding association rules to obtain multiple association rule sets. Filter the first frequent items according to the association rule sets to obtain the event combination set. The association rules in the association rule set are used to characterize the causal relationship between the above-mentioned public opinion event combinations and the above-mentioned first preset risk event.

[0121] Specifically, this involves grasping the evolution patterns of minor public opinion events and predicting the next major risk event based on historical minor public opinion events. Based on the current set of public opinion events and an initial support threshold (preset support), the Apriori algorithm is used to discover frequent itemsets from a large number of event labels—that is, combinations of minor public opinion events that may lead to major risk events. Furthermore, association rules are constructed based on the extracted combinations to clarify their risk orientation and to further filter combinations with strong correlations with risk events.

[0122] Understandably, a single public opinion event may not be enough to cause significant fluctuations in corporate bond prices, but a combination of events can create substantial risks. By constructing frequent itemsets and corresponding association rule sets, the causal relationships between events can be quantified, and the risks that event combinations may trigger can be identified. This allows for more accurate prediction of corporate bond price fluctuations, solving the problem of insufficient prediction for isolated events.

[0123] Through the above embodiments, by filtering frequent itemsets and constructing association rules, causal relationships between events are discovered, thereby predicting abnormal price movements in credit bonds. This method utilizes hidden correlations between data, enhancing the interpretability and reliability of the predictive model.

[0124] To filter frequent itemsets, in one optional implementation, step S20331 above includes:

[0125] Step S203311: Using a single public opinion event as the basic element, convert all public opinion events in the updated wide table into the basic elements in the dataset to obtain the candidate frequent itemset.

[0126] Specifically, at the beginning of the algorithm, the first level (which can be set to C1) of the candidate frequent itemsets is constructed from the original public opinion data. That is, the above-mentioned wide table of updated data is converted into a set with a single public opinion event as the basic unit.

[0127] Step S203312: Calculate the support of each basic element in the above-mentioned candidate frequent itemset based on the above-mentioned updated data wide table to obtain the first support.

[0128] Specifically, the support of each basic element is calculated by traversing the above-mentioned candidate frequent itemsets to obtain the above-mentioned first support. It can be understood that the support depends on the frequency of the occurrence of the public opinion event or combination of public opinion events in the overall data.

[0129] Step S203313: Construct a second frequent itemset based on the basic elements whose first support is greater than the preset support.

[0130] Specifically, all basic elements with support greater than the preset support are placed into frequent itemsets to obtain the first level of frequent itemsets (which can be set to L1). The support threshold α is a predefined parameter used to determine whether an itemset is considered frequent. It depends on the depth of the event correlation analysis. The higher the depth, the lower the support threshold, but the difficulty of subsequent analysis and calculation will also increase.

[0131] Step S203314, the first update step, uses the random combination of the basic elements in the second frequent itemset as the basic elements to update the candidate frequent itemset, and obtains the updated candidate frequent itemset. In the updated candidate frequent itemset, any two basic elements are inconsistent.

[0132] Specifically, a second level (C2) of candidate frequent itemsets is constructed based on the first level of frequent itemsets. In other words, the basic elements in L1 are traversed and combined to cover all possible combinations, resulting in an updated candidate frequent itemset.

[0133] Step S203315, the second update step, recalculate the support based on the updated candidate frequent itemset to obtain the second support, and update the second frequent itemset according to the second support;

[0134] Specifically, based on the second level of the candidate frequent itemsets, the second level (L2) of the frequent itemsets is filtered. Specifically, all basic elements with support greater than the preset support are put into the frequent itemsets to obtain the updated second frequent itemsets.

[0135] Step S203316: Repeat the first update step and the second update step in sequence until the second support corresponding to the basic elements in the updated candidate frequent itemset is less than the preset support, and determine the last updated second frequent itemset as the first frequent itemset.

[0136] Specifically, the candidate frequent itemset and higher levels of the frequent itemset are repeatedly constructed until no higher levels of the frequent itemset can be generated, indicating that the frequent itemset has already included all possible combinations of minor public opinion events that could lead to risk events.

[0137] Understandably, constructing frequent itemsets is a crucial step in uncovering causal relationships between event combinations. Iteratively updating these frequent itemsets ensures that the selected event combinations indeed possess high support, reflecting the true patterns in the public opinion data. This selection mechanism helps reduce the computational burden of model training while improving the quality and accuracy of the predictive model.

[0138] Through the above embodiments, itemsets with high support are gradually selected through iterative updates, providing a foundation for constructing event combination sets. This method ensures that the selected event combination sets are based on a true reflection of historical data, improving the accuracy of predictions.

[0139] In order to construct the association rule set, in one optional implementation, step S20332 above includes:

[0140] Step S203321: Based on each of the basic elements in the first frequent itemset, establish the association rules with each of the first preset risk events, wherein the association rules are based on the basic elements and the first preset risk events.

[0141] Step S203322: Calculate the corresponding confidence level based on each of the above association rules to obtain the target confidence level;

[0142] Step S203323: If the target confidence level is greater than the preset confidence level, calculate the lift of the association rule based on the target confidence level to obtain the target lift.

[0143] Step S203324: Construct the above association rule set based on the above association rules whose target lift is greater than the preset lift;

[0144] Step S203325: Update the first frequent itemset with the basic elements included in the association rule set, and merge the updated first frequent itemset and the association rule set to obtain the event combination set.

[0145] Specifically, after generating a series of frequent itemsets using the Apriori algorithm, rules for A and B are generated starting from L1, where A is the premise and B is the conclusion. In this application, the aforementioned basic elements are used as premises, and the aforementioned first preset risk event is used as the conclusion. The generated rules are then filtered based on confidence levels, where confidence depends on the probability of occurrence given A. Conditional probability quantifies the probability of a risk event occurring after a minor public opinion event. The filtered rules are further filtered using lift, where lift is defined as the product of the AB confidence level and the independent support of A and B. A lift greater than 1 indicates that the probability of B occurring after A is much greater than the probability of A occurring independently, meaning there is a strong causal relationship between A and B. Conversely, a lift less than 1 indicates a weak causal relationship between A and B. The frequent itemsets are then further filtered based on the filtered rules to remove poorly performing data, resulting in the aforementioned event combination set.

[0146] Understandably, the rules in the association rule set reflect the strong correlation between the event combination and the preset risk events. The calculation of confidence and lift can quantify this correlation, ensuring that the constructed rule set is effective. By setting confidence and lift thresholds, truly valuable associations can be selected, which can improve the reliability and practicality of credit bond risk prediction and avoid the analytical difficulty and prediction errors caused by an overly large rule set.

[0147] Through the above embodiments, by calculating confidence and lift, combinations of events highly correlated with preset risk events are selected. This method can quantify the strength of the causal relationship between events, helping to more accurately identify potential risk signals.

[0148] In order to determine the event tag for each public opinion event, in one optional implementation, step S201 above includes:

[0149] Step S2011: Retrieve news data of the aforementioned bond issuers from multiple preset paths to obtain the fifth public opinion data; perform data cleaning on the aforementioned fifth public opinion data to obtain the aforementioned first public opinion data.

[0150] Step S2012: Based on the first preset risk events existing in the business scenarios involved in the first public opinion data, determine the category to which the first public opinion data belongs, and obtain a first-level label; and determine the category of the public opinion events that caused the first preset risk events to which the first public opinion data belongs, and obtain a second-level label. The first-level label includes financial risk, operational risk, management risk and legal risk; the second-level label includes payment delay, product quality inspection issues, management personnel changes, litigation events and supply chain changes.

[0151] Step S2013: The above-mentioned primary tags and secondary tags are determined as the above-mentioned first event tags of the above-mentioned first public opinion data.

[0152] Understandably, accurate classification and labeling of public opinion events are fundamental to risk prediction. By setting primary and secondary labels, public opinion data can be systematically organized and understood, ensuring that the model can capture different types of credit risk. This classification method improves the efficiency and accuracy of event identification, which is helpful for subsequent risk quantification and the construction of predictive models.

[0153] By capturing and preprocessing public opinion data from multiple information sources, a comprehensive event tagging system was constructed, solving the problems of incomplete data acquisition and non-standard preprocessing. Through the establishment of primary and secondary tags, this solution can classify public opinion events, identify event types that have a potential impact on credit bond price fluctuations, and provide clear guidance for subsequent event analysis and risk prediction. The use of event tags not only improves the model's predictive accuracy but also reveals the specific impact path of events on credit bond price fluctuations, providing financial institutions and investors with in-depth risk insights.

[0154] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0155] This application also provides a device for predicting credit bond price fluctuations based on public opinion analysis. It should be noted that this device can be used to execute the method for predicting credit bond price fluctuations based on public opinion analysis provided in this application. This device is used to implement the above embodiments and preferred embodiments; details already described will not be repeated. As used below, the term "module" can refer to a combination of software and / or hardware that performs a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0156] The following describes the credit bond price fluctuation prediction device based on public opinion analysis provided in the embodiments of this application.

[0157] Figure 3 This is a structural block diagram of a credit bond price fluctuation prediction device based on public opinion analysis, according to an embodiment of this application. Figure 3 As shown, the device includes:

[0158] The first acquisition unit 10 is used to acquire news data of the bond issuer to obtain first public opinion data, acquire news data of related entities of the bond issuer to obtain second public opinion data, determine public opinion events of the bond issuer based on the first public opinion data to obtain first event tags, determine the aforementioned public opinion events of the related entities based on the second public opinion data to obtain second event tags, and determine the risk transmission coefficient based on the association between the bond issuer and the related entities. The risk transmission coefficient is used to characterize the degree of influence between the bond issuer and the related entities.

[0159] The first processing unit 20 is used to associate the second event label and the second public opinion data with the first event label and the first public opinion data according to the risk transmission coefficient, so as to obtain a wide data table;

[0160] The second processing unit 30 is used to process the above-mentioned wide data table using the APRIORI algorithm to obtain an event combination set, which includes at least a combination of public opinion events with risk orientation.

[0161] The third processing unit 40 is used to process the above-mentioned wide data table through the LSTM algorithm to obtain an event chain. The event chain is used to characterize the change trend of the above-mentioned public opinion event within a preset first time period after the current moment.

[0162] The prediction unit 50 is used to predict the probability of price fluctuation of the bond issuer based on the above event combination set and the above event chain using a logistic regression model, or to predict the probability of the occurrence of the first preset risk event based on the above event chain using an MLP model. When the above price fluctuation probability or the above risk probability is greater than the corresponding preset threshold, a target alarm message is generated. The above logistic regression model is trained based on the above public opinion event combination, the changing trend of the above public opinion events and the fluctuation of credit bond prices in the historical period. The above MLP model is trained based on whether the above event chain and the above first preset risk event have occurred in the historical period.

[0163] This embodiment integrates public opinion data by mining the upstream and downstream relationships between the bond issuer and related entities. Based on this integrated data, the APRIORI algorithm is used to mine the correlations between public opinion data, and the LSTM algorithm predicts the development trend of public opinion events over time to uncover potentially risky public opinion data. A logistic regression model is used to synthesize the correlations between public opinion data and the current trend of public opinion data to predict future credit bond price fluctuations that may be caused by changes in public opinion, thus avoiding losses. Compared to existing technologies that rely on the proportion of key public opinion and negative public opinion, this application improves upon existing technologies by expanding the dimensions of data analysis, introducing the mutual influence between public opinion data of different entities, mining the correlations between public opinion data, introducing the mutual influence between different public opinion events, and mining the development chain of public opinion events to avoid missing potential public opinion. Furthermore, by using a logistic regression model to synthesize the current public opinion situation resulting from the influence of different events and predicting the development trend based on the current public opinion events, this invention solves the problem that existing credit bond assessment methods lack sufficient public opinion data mining, leading to low accuracy in assessment results and causing financial losses for enterprises.

[0164] To integrate the aforementioned first and second public opinion data, in one optional implementation, the first processing unit includes:

[0165] The first processing module is used to perform word segmentation on the first public opinion data using NLP technology, and to remove linking words, stop words and part-of-speech tags from the first public opinion data after word segmentation to obtain the third public opinion data. The module also performs word segmentation on the second public opinion data using NLP technology, and to remove linking words, stop words and part-of-speech tags from the second public opinion data after word segmentation to obtain the fourth public opinion data.

[0166] The second processing module is used to convert the aforementioned third public opinion data into a feature vector using the TF-IDF algorithm to obtain a first target vector, and to convert the aforementioned fourth public opinion data into a feature vector using the TF-IDF algorithm to obtain a second target vector.

[0167] The first clustering module is used to cluster the first target vector to obtain multiple first target clusters, and the first target clusters correspond one-to-one with the first event labels.

[0168] The second clustering module is used to cluster the second target vector to obtain multiple second target clusters, and the second target clusters correspond one-to-one with the second event labels.

[0169] The third processing module is used to aggregate the second target cluster with the corresponding first target cluster based on the first event label, the second event label and the risk transmission coefficient to obtain multiple third target clusters, and to determine the corresponding event label according to each third target cluster to obtain the target event label.

[0170] The module is used to fill in the third target cluster with the target event label as the table header to obtain the wide data table.

[0171] To analyze the public opinion situation, in one optional implementation, the second processing unit includes:

[0172] The first determining module is used to determine the risk level of the aforementioned public opinion event based on the aforementioned third target cluster corresponding to each of the aforementioned target event tags, and to obtain the target risk level.

[0173] The update module is used to filter out the third target clusters in the above data wide table that correspond to the target risk level greater than the preset level, and obtain the updated data wide table;

[0174] The fourth processing module is used to process the updated wide data table using the APRIORI algorithm to obtain the event combination set.

[0175] To uncover the aggregation effect of subtle public opinion events, in one optional implementation, the fourth processing module includes:

[0176] The filtering submodule is used to filter combinations of public opinion events with a support greater than the preset support at different scales based on the above updated data wide table, and obtain the first frequent itemset;

[0177] A submodule is constructed to build corresponding association rules based on each of the above-mentioned combinations of public opinion events in the first frequent item set, thereby obtaining multiple association rule sets. The first frequent items are then filtered according to the association rule sets to obtain the event combination set. The association rules in the association rule set are used to characterize the causal relationship between the combination of public opinion events and the first preset risk event.

[0178] To filter frequent itemsets, in one optional implementation, the filtering module includes:

[0179] The first processing submodule is used to convert all the public opinion events in the above-mentioned updated data wide table into the above-mentioned basic elements in the dataset, using a single public opinion event as the basic element, to obtain a candidate frequent itemset.

[0180] The first calculation submodule is used to calculate the support of each of the basic elements in the above-mentioned candidate frequent itemset based on the above-mentioned updated data wide table, and obtain the first support.

[0181] The second processing submodule is used to construct a second frequent itemset based on the basic elements whose first support is greater than the preset support.

[0182] The first update submodule is used to perform the first update step, which uses the basic elements in the second frequent itemset to be randomly combined in pairs as the basic elements, updates the candidate frequent itemset, and obtains the updated candidate frequent itemset. Any two basic elements in the updated candidate frequent itemset are inconsistent.

[0183] The second update submodule is used to perform the second update step, recalculate the support based on the updated candidate frequent itemset to obtain the second support, and update the second frequent itemset according to the second support.

[0184] The repeating submodule is used to repeat the first update step and the second update step in sequence until the second support corresponding to the basic elements in the updated candidate frequent itemset is less than the preset support, and the last updated second frequent itemset is determined as the first frequent itemset.

[0185] To construct the association rule set, in one optional implementation, the above-mentioned construction submodule includes:

[0186] The third processing submodule is used to establish the association rules based on each of the basic elements in the first frequent itemset and each of the first preset risk events, wherein the association rules are based on the basic elements and the first preset risk events.

[0187] The second calculation submodule is used to calculate the corresponding confidence scores based on each of the above association rules to obtain the target confidence score.

[0188] The third calculation submodule is used to calculate the lift of the association rule based on the target confidence level when the target confidence level is greater than the preset confidence level, and obtain the target lift.

[0189] The fourth processing submodule is used to construct the above association rule set based on the above association rules whose target lift is greater than the preset lift;

[0190] The third update submodule is used to update the first frequent itemset with the basic elements included in the association rule set, and merge the updated first frequent itemset and the association rule set to obtain the event combination set.

[0191] To determine the event tags for each public opinion event, in one optional implementation, the first acquisition unit includes:

[0192] The first acquisition module is used to capture news data of the aforementioned bond issuers from multiple preset paths to obtain the fifth public opinion data, and to perform data cleaning on the aforementioned fifth public opinion data to obtain the aforementioned first public opinion data.

[0193] The second determining module is used to determine the category to which the first public opinion data belongs based on the first preset risk event existing in the business scenario involved by the first public opinion data, to obtain a first-level label, and to determine the category of the public opinion event to which the first public opinion data belongs that caused the first preset risk event, to obtain a second-level label. The first-level label includes financial risk, operational risk, management risk and legal risk, and the second-level label includes payment delay, product quality inspection issues, management personnel changes, litigation events and supply chain changes.

[0194] The third determining module is used to determine the above-mentioned first-level tags and the above-mentioned second-level tags as the above-mentioned first event tags of the above-mentioned first public opinion data.

[0195] The aforementioned credit bond price fluctuation prediction device based on public opinion analysis includes a processor and a memory. The first acquisition unit, first processing unit, second processing unit, third processing unit, and prediction unit are all stored as program units in the memory. The processor executes these program units stored in the memory to achieve the corresponding functions. All of the above modules are located in the same processor; alternatively, the modules may be located in different processors in any combination.

[0196] The processor contains a kernel, which retrieves the corresponding program units from memory. One or more kernels can be configured, and adjusting kernel parameters can improve the accuracy of credit bond risk prediction.

[0197] The memory may include non-permanent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.

[0198] This invention provides a computer-readable storage medium including a stored program, wherein, when the program is executed, it controls the device containing the computer-readable storage medium to execute the aforementioned credit bond price fluctuation prediction method based on public opinion analysis.

[0199] This invention provides a processor for running a program, wherein the program executes the aforementioned credit bond price fluctuation prediction method based on public opinion analysis.

[0200] This invention provides a risk monitoring system, which includes a processor, a memory, and a program stored in the memory and executable on the processor. When the processor executes the program, it implements at least the steps of a credit bond price fluctuation prediction method based on public opinion analysis.

[0201] This application also provides a computer program product that, when executed on a data processing device, is adapted to perform a program that initializes a method for predicting credit bond price fluctuations based at least on public opinion analysis.

[0202] It is obvious to those skilled in the art that the modules or steps of the present invention described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. They can be implemented using computer-executable program code, and thus can be stored in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those described herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the present invention is not limited to any particular combination of hardware and software.

[0203] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0204] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0205] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0206] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0207] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0208] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0209] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0210] It should also be noted that the terms "comprising," "including," or any other variations thereof are 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 process, method, article, or apparatus. Unless otherwise specified, 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 that element.

[0211] As can be seen from the above description, the embodiments of this application achieve the following technical effects:

[0212] 1) This application's method for predicting credit bond price fluctuations based on public opinion analysis integrates public opinion data by mining the upstream and downstream relationships between bond issuers and related entities. Then, based on the integrated public opinion data, it uses APRIORI to mine the correlations between public opinion data, and the LSTM algorithm predicts the development trend of public opinion events from a timeline perspective to uncover potentially risky public opinion data. A logistic regression model is used to synthesize the correlations between public opinion data and the current development trend of public opinion data to predict future credit bond price anomalies caused by changes in public opinion, thus avoiding losses. Compared to existing technologies that rely on the proportion of key public opinion and negative public opinion, this application improves upon existing technologies by expanding the dimensions of data analysis, introducing the mutual influence between public opinion data of different entities, mining the correlations between public opinion data, introducing the mutual influence between different public opinion events, and mining the development chain of public opinion events to avoid missing potential public opinion. Furthermore, by using a logistic regression model to synthesize the current public opinion situation resulting from the influence of different events and predicting the development trend based on the current public opinion events, this application solves the problem of insufficient public opinion data mining in existing credit bond assessment methods, leading to low accuracy in assessment results and causing financial losses for enterprises.

[0213] 2) This application's credit bond price fluctuation prediction device based on public opinion analysis integrates public opinion data by mining the upstream and downstream relationships between bond issuers and related entities. Based on the integrated public opinion data, it uses APRIORI to mine the correlations between public opinion data, and the LSTM algorithm predicts the development trend of public opinion events from a timeline perspective to uncover potentially risky public opinion data. A logistic regression model is used to synthesize the correlations between public opinion data and the current development trend of public opinion data to predict future credit bond price anomalies caused by changes in public opinion, thus avoiding losses. Compared to existing technologies that rely on the proportion of key public opinion and negative public opinion, this application improves upon existing technologies by expanding the dimensions of data analysis, introducing the mutual influence between public opinion data of different entities, mining the correlations between public opinion data, introducing the mutual influence between different public opinion events, and mining the development chain of public opinion events to avoid missing potential public opinion. Furthermore, by using a logistic regression model to synthesize the current public opinion situation resulting from the influence of different events and predicting the development trend based on the current public opinion events, this application solves the problem of insufficient public opinion data mining in existing credit bond assessment methods, leading to low accuracy in assessment results and causing financial losses for enterprises.

[0214] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for predicting credit bond price fluctuations based on public opinion analysis, characterized in that, include: The process involves acquiring news data of the bond issuer to obtain first public opinion data, acquiring news data of related entities of the bond issuer to obtain second public opinion data, determining public opinion events of the bond issuer based on the first public opinion data to obtain a first event tag, determining the public opinion events of the related entities based on the second public opinion data to obtain a second event tag, and determining a risk transmission coefficient based on the association between the bond issuer and the related entities. The risk transmission coefficient is used to characterize the degree of influence between the bond issuer and the related entities. The second event tag and the second public opinion data are associated with the first event tag and the first public opinion data based on the risk transmission coefficient to obtain a wide data table; The data wide table is processed by the APRIORI algorithm to obtain an event combination set, which includes at least a combination of public opinion events with risk orientation. The data wide table is processed by the LSTM algorithm to obtain an event chain, which is used to characterize the changing trend of the public opinion event within a preset first time period after the current moment. Based on the event set and the event chain, a logistic regression model is used to predict the probability of price fluctuations of the bond issuer, or an MLP model is used to predict the probability of a first preset risk event occurring. If the probability of price fluctuations or the probability of risk exceeds a corresponding preset threshold, a target alarm message is generated. The logistic regression model is trained based on the combination of public opinion events, the changing trend of the public opinion events, and the fluctuation of credit bond prices in historical periods. The MLP model is trained based on whether the event chain and the first preset risk event have occurred in historical periods.

2. The method according to claim 1, characterized in that, Based on the risk transmission coefficient, the second event tag and the second public opinion data are associated with the first event tag and the first public opinion data to obtain a wide data table, including: The first public opinion data is segmented using NLP technology, and the connecting words, stop words, and part-of-speech tags in the segmented first public opinion data are removed to obtain the third public opinion data. The second public opinion data is segmented using the same NLP technology, and the connecting words, stop words, and part-of-speech tags in the segmented second public opinion data are removed to obtain the fourth public opinion data. The third public opinion data is converted into a feature vector using the TF-IDF algorithm to obtain the first target vector. The fourth public opinion data is then converted into a feature vector using the same TF-IDF algorithm to obtain the second target vector. The first target vector is clustered to obtain multiple first target clusters, and each first target cluster corresponds one-to-one with the first event label; The second target vector is clustered to obtain multiple second target clusters, and each second target cluster corresponds one-to-one with the second event label; Based on the first event label, the second event label, and the risk transmission coefficient, the second target cluster is aggregated with the corresponding first target cluster to obtain multiple third target clusters. The corresponding event label is determined according to each third target cluster to obtain the target event label. The third target cluster is filled in with the target event label as the table header to obtain the wide data table.

3. The method according to claim 2, characterized in that, The data wide table is processed using the APRIORI algorithm to obtain an event combination set, including: The risk level of the corresponding public opinion event is determined based on the third target cluster corresponding to each target event label, thus obtaining the target risk level. The third target clusters in the data wide table that correspond to a target risk level greater than a preset level are filtered out to obtain an updated data wide table; The event combination set is obtained by processing the updated wide data table using the APRIORI algorithm.

4. The method according to claim 3, characterized in that, The updated data wide table is processed using the APRIORI algorithm to obtain the event combination set, including: Based on the updated wide data table, combinations of public opinion events with support greater than preset support are filtered at different scales to obtain the first frequent itemset; Based on each of the public opinion event combinations in the first frequent item set, corresponding association rules are constructed to obtain multiple association rule sets. The first frequent items are filtered according to the association rule sets to obtain the event combination set. The association rules in the association rule set are used to characterize the causal relationship between the combination of public opinion events and the first preset risk event.

5. The method according to claim 4, characterized in that, Based on the updated wide data table, combinations of public opinion events with support greater than a preset support are filtered at different scales to obtain a first frequent itemset, including: Using a single public opinion event as the basic element, all public opinion events in the updated data wide table are converted into the basic element in the dataset to obtain a candidate frequent itemset; The support of each basic element in the candidate frequent itemset is calculated based on the updated data wide table to obtain the first support. Construct a second frequent itemset based on the basic elements whose first support is greater than the preset support; The first update step involves using random combinations of the basic elements in the second frequent itemset as the basic elements to update the candidate frequent itemset, resulting in the updated candidate frequent itemset. In the updated candidate frequent itemset, any two basic elements are inconsistent. The second update step involves recalculating the support based on the updated candidate frequent itemset to obtain a second support, and updating the second frequent itemset according to the second support. The first update step and the second update step are repeated sequentially until the second support corresponding to the basic element in the updated candidate frequent itemset is less than the preset support. The second frequent itemset updated last time is then determined as the first frequent itemset.

6. The method according to claim 5, characterized in that, Based on each of the aforementioned public opinion event combinations in the first frequent item set, corresponding association rules are constructed to obtain multiple association rule sets. The first frequent items are then filtered according to these association rule sets to obtain the event combination set, including: The association rules are established based on each of the basic elements in the first frequent itemset and each of the first preset risk events, wherein the association rules are based on the basic elements and the first preset risk events are the conclusions. The target confidence level is obtained by calculating the corresponding confidence level based on each of the association rules. If the target confidence level is greater than the preset confidence level, the lift of the association rule is calculated based on the target confidence level to obtain the target lift. The association rule set is constructed based on the association rules whose target lift is greater than a preset lift. The first frequent itemset is updated with the basic elements included in the association rule set, and the updated first frequent itemset and the association rule set are merged to obtain the event combination set.

7. The method according to claim 1, characterized in that, Obtain news data of the bond issuer to obtain primary public opinion data. Based on the primary public opinion data, determine the public opinion events related to the bond issuer and obtain primary event tags, including: News data of the bond issuer is captured from multiple preset paths to obtain the fifth public opinion data. The fifth public opinion data is then cleaned to obtain the first public opinion data. Based on the first preset risk event existing in the business scenario involved by the first public opinion data, the category to which the first public opinion data belongs is determined to obtain a first-level label, and the category of the public opinion event to which the first public opinion data belongs that caused the first preset risk event is determined to obtain a second-level label. The first-level label includes financial risk, operational risk, management risk and legal risk, and the second-level label includes payment delay, product quality inspection issues, management personnel changes, litigation events and supply chain changes. The primary label and the secondary label are determined as the first event label of the first public opinion data.

8. A device for predicting credit bond price fluctuations based on public opinion analysis, characterized in that, The device includes: The first acquisition unit is used to acquire news data of the bond issuer to obtain first public opinion data, acquire news data of related entities to obtain second public opinion data, determine the public opinion event of the bond issuer based on the first public opinion data to obtain a first event tag, determine the public opinion event of the related entities based on the second public opinion data to obtain a second event tag, and determine the risk transmission coefficient based on the association between the bond issuer and the related entities. The risk transmission coefficient is used to characterize the degree of influence between the bond issuer and the related entities. The first processing unit is used to associate the second event tag and the second public opinion data with the first event tag and the first public opinion data according to the risk transmission coefficient, so as to obtain a wide data table; The second processing unit is used to process the wide data table using the APRIORI algorithm to obtain an event combination set, wherein the event combination set includes at least a combination of public opinion events with risk orientation. The third processing unit is used to process the wide data table using the LSTM algorithm to obtain an event chain, which is used to characterize the trend of the public opinion event within a preset first time period after the current moment. The prediction unit is used to predict the probability of price fluctuation of the bond issuer based on the event combination set and the event chain using a logistic regression model. If the probability of price fluctuation is greater than a preset threshold, a target alarm message is generated. The logistic regression model is trained based on the combination of public opinion events in historical periods, the changing trend of the public opinion events, and the fluctuation of credit bond prices.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored program, wherein, when the program is executed, it controls the device on which the computer-readable storage medium is located to perform the method according to any one of claims 1 to 7.

10. A risk monitoring system, characterized in that, include: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs comprising methods for performing any one of claims 1 to 7.

Images