Stock price collapse risk analysis reference system based on stock market interaction data identification

Abstract

The invention discloses a stock price collapse risk analysis reference system based on stock market interaction data identification, and relates to the technical field of stock market management systems. The stock price collapse risk analysis reference system based on stock market interaction data identification comprises a multi-modal data acquisition and processing module, a group behavior mode modeling analysis module, an enhanced emotion analysis and antagonism defense module and a risk early warning model construction and verification module. By integrating multi-source data and applying an attention mechanism to dynamically distribute weights, the comprehensiveness and accuracy of data processing are improved; a directed weighted graph and an SIR model are utilized, key nodes are identified more accurately, and negative emotion diffusion is simulated; an XGBoost model is adopted to output an emotion label, so that the accuracy of emotion recognition is enhanced; noise interference is effectively filtered by generating an adversarial sample and a dual machine learning framework; multiple factors are integrated, the weight is dynamically adjusted by adopting a LightGBM model, and the stability of the model and the accuracy of early warning are ensured through back testing and real-time monitoring.

Description

Technical Field

[0001] This invention relates to the field of technology, specifically to a stock price crash risk analysis reference system based on stock market interaction data identification. Background Technology

[0003] Publication No. CN113240346A discloses a method for analyzing the impact of accounting firm reputation on stock price crash risk, including: a sample data acquisition step, which processes the sample source data to obtain sample data; a variable acquisition step, which obtains the dependent variable data, explanatory variable data, and control variable data based on the sample data; the explanatory variables are metrics related to accounting firm reputation; a model parameter estimation step, which estimates the parameters based on the sample data to obtain parameter estimates; an analytical model construction step, which constructs an analytical model based on the dependent variable, explanatory variables, and control variables; and a predictive analysis step, which predicts and analyzes the stock price crash risk based on the analytical model to obtain the stock price crash risk analysis results. This method enriches the research system of factors influencing stock price crash risk from a new perspective.

[0004] As shown in the aforementioned technologies, current analysis and prediction of stock price crash risks rely on financial data analysis, using the changing patterns of stock market financial data to predict crash risks. However, this requires strong professional knowledge and rich experience, which most stock market investors lack. They mostly blindly follow the crowd, leading to significant losses during stock market crashes. Therefore, this application proposes a stock price crash risk analysis reference system based on stock market interaction data identification. By analyzing the interaction data of other investors on the stock market interaction platform, this system analyzes the tendencies (such as emotional changes) of other experienced investors to provide inexperienced investors with a channel for predicting stock price crash risks, thereby reducing investor losses. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a stock price crash risk analysis reference system based on stock market interaction data identification, thus solving the aforementioned problems in stock market investment.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a stock price crash risk analysis reference system based on stock market interaction data identification, comprising: Multimodal data acquisition and processing module: Collects text, charts, emoticons, and voice data of investor interactions from online platforms, and simultaneously obtains stock market data and transaction data; extracts key price levels and fund flows from charts, performs text-chart semantic matching, performs sentiment analysis and enhancement on emoticons and voice respectively, synchronizes data by timestamp, and dynamically allocates weights using an attention mechanism; Group behavior pattern modeling and analysis module: Constructs a directed weighted graph based on user relationships, identifies key nodes and marks emotional mutation points; clusters post topics, calculates semantic drift and topic negativity index, defines topic escalation rules; constructs an emotion propagation SIR model to simulate the speed and scope of negative emotion spread; Enhanced sentiment analysis and adversarial defense module: Combining multimodal data and group behavior characteristics, constructing sentiment feature vectors, outputting sentiment classification labels and confidence scores through the XGBoost model; generating adversarial examples to evaluate model robustness, optimizing the loss function using an adversarial training strategy, and estimating the average treatment effect based on a dual machine learning framework, constructing counterfactual scenarios to filter noise interference; Risk warning model construction and verification module: Integrates sentiment, market, and capital factors, uses the LightGBM model to train a risk classifier, and dynamically adjusts factor weights; ensures model stability and accuracy through backtesting, real-time monitoring, and stress testing.

[0007] Preferably, the multimodal data acquisition and processing module acquires data including: text J, charts C, emoticons E, and voice V from investor interaction platforms and financial video platforms, and simultaneously obtains real-time stock quotes M and historical transaction data L for the corresponding stocks.

[0008] Preferably, the data processing steps performed by the multimodal data acquisition and processing module include: Chart processing: Using OCR technology, the support level (S) and resistance level (R) values ​​in the candlestick chart, and the net inflow (I) and net outflow (O) amounts in the fund flow chart, are extracted to construct a chart feature vector. I O>0 This is an indicator function that takes the value 1 when 0 > 0, and 0 otherwise. Text-Chart Semantic Matching: Perform semantic matching verification between the price point P mentioned in the text and the key price point on the chart, and calculate the Euclidean distance. If d is less than the threshold d', it is marked as consistent; otherwise, manual review is triggered. Here, k is a variable that takes a value from the set {S, R}. Emoji Sentiment Analysis: A pre-trained visual ransformer model is used to classify emojis, outputting emotion labels such as anger, panic, and optimism, along with probability distributions e=[e1, e2, ..., e...]. K A comprehensive sentiment index is generated by combining the text sentiment weight w: Where α is the emotion weight of the emoji and β is the text-emoji fusion coefficient; Speech emotion enhancement: An end-to-end ASR model is used to transcribe speech content, while simultaneously extracting Mel-frequency cepstral coefficient features. An LSTM network is then used to predict the speech rate. t Pitch p t Energy et Voiceprint feature parameters are then mapped to emotion intensity correction factors, specifically including: concatenating the features into a vector. The input layer of the LSTM network is a sequence {x1, x2, ..., x}. T}, where T is the time step; the LSTM layer is used for hidden state updates: The output layer calculates the correction factor through a fully connected layer. Among them, W s This represents the weight matrix of the output layer, which maps the hidden states to correction factors; b s Here, σ is the bias term of the output layer; σ is the σ function, which compresses the output to the (0, 1) interval or directly uses linear activation; the final emotion intensity is... ; Data alignment and fusion: Synchronize multimodal data by timestamp and dynamically allocate modality weights using an attention mechanism. , where q t For the query vector, p i v i These are the key-value pairs for each modality.

[0009] Preferably, the steps of the group behavior pattern modeling and analysis module to analyze group behavior patterns include social network construction, topic evolution tracking, and quantification of group emotion propagation. The steps of social network construction include: constructing a directed weighted graph based on user ID and reply / forward relationship; node attributes include user historical sentiment tendency and influence score; user historical sentiment tendency is represented by the ratio of positive to negative posts; influence score is represented by PageRank value; edge weights are interaction frequency and sentiment consistency; the PageRank value calculation formula is: Where N represents the total number of nodes in the social network, i.e. the number of users, d is the damping coefficient, which takes a value of 0-1, and In(u) and Out(v) are the sets of incoming links for user u and outgoing links for user v, respectively. Identify key moments: Select the top 10% of users by PageRank as KOLs and mark their emotional turning points: This is used to calculate the degree of emotional change ΔE of user u at time t. u (t), and determine whether the change exceeds a statistical threshold based on the sentiment values ​​of the past 24 hours, where mean is the mean function. Indicates user u is from The sentiment value sequence up to t over the past 24 hours; std is the standard deviation function; when When this occurs, it indicates that the current emotional value at time t deviates significantly from the average emotional level of the past 24 hours, meaning that user u experienced a sudden emotional shift at time t. In this case, ΔE... uIf (t) is true, it is marked as an emotional mutation point; otherwise, it is not considered that a significant emotional mutation has occurred.

[0010] Preferably, the topic evolution tracking steps include: clustering post topics using BERTopic by hourly time windows, generating topic word vectors and calculating semantic drift, and calculating a topic negativity index using a sentiment dictionary. ;where P t Let be the set of keywords, neg(f) be the weight of negative words, and λ be the weight of negative words. t As an emotion intensity attenuation factor, η=0.1, Δt is the time interval; Define topic escalation rules: When the negative index of a topic rises for two consecutive hours and the turnover rate of related stocks increases simultaneously, it is marked as a high-risk topic, triggering a backtracking of the propagation path.

[0011] Preferably, the step of quantifying the spread of group emotions includes: constructing an emotion spread SIR model, classifying users into three categories: uninvolved (A1), affected (A2), and immune (A3), dynamically simulating the spread speed and coverage of negative emotions in social networks, and predicting the arrival time of the emotion peak. The dynamic equation of the SIR model is: ;in For the transmission rate, The recovery rate is obtained by fitting historical data.

[0012] Preferably, the enhanced sentiment analysis and adversarial defense module includes the following for enhanced sentiment analysis: constructing a sentiment feature vector by combining multimodal data and group behavioral characteristics: Among them, NER t λ is the word frequency statistical vector of the text named entity recognition results. KOL This refers to the propagation coefficient of key nodes, which is the weight after normalization based on PageRank values; Output sentiment classification labels using the XGBoost model: ; It also outputs the confidence score, calculated using Softmax: c∈{positive, neutral, negative}; where ϕ c Let j be the raw score of category c output by XGBoost, and j be an index variable for summation.

[0013] Preferably, the enhanced sentiment analysis and adversarial defense module includes the following adversarial defense features: Adversarial text processing: TextFooler is used to generate adversarial examples, constructing a test set containing implicit expressions to evaluate model robustness; the adversarial training strategy is as follows: perturbation text is added during the model fine-tuning stage, adversarial noise is generated using FGM, and the loss function is optimized as a weighted sum of the original loss and the adversarial loss: L total =(1−b)L original +bL adv ; where L original L represents the original loss value calculated based on normal training data. adv denoted as the adversarial loss value calculated after adding perturbation text, where b is the proportion of perturbation samples; Counterfactual causal reasoning: Based on the dual machine learning DML framework, sentiment text is used as the processing variable and stock price change is used as the outcome variable. After controlling for market factors, the average processing effect is estimated. Counterfactual scenarios are constructed: the stock price path under the condition that a negative sentiment text has not been published is simulated by generative adversarial network GAN. The difference between the actual path and the counterfactual path is compared to filter out noise interference.

[0014] Preferably, the risk warning model construction and verification module constructs the model by: integrating sentiment factors, market factors, and capital factors to construct a feature matrix F, and training a risk classifier using the LightGBM model; the feature matrix F is represented as: Among them, NegSent represents the speed of negative sentiment spread, reflecting the extent of negative sentiment diffusion in the market, and belongs to the sentiment factor; TopicNeg represents the topic negativity index, measuring the intensity of negative sentiment for a specific topic, and belongs to the sentiment factor; VIX is the market volatility index, reflecting market uncertainty, and belongs to the market factor; Turnover represents sudden changes in turnover rate, reflecting changes in market trading activity, and belongs to the market factor; DiscountRate represents the discount rate for large transactions, reflecting the discount situation of large transactions, indirectly reflecting the tightness of the funding situation, and belongs to the funding factor; Dynamic weight adjustment: Automatically adjusts factor weights based on market conditions; assuming three market conditions: bull market, bear market, and sideways market; automatically adjusts the weights of each factor B according to different market conditions. ; Warning threshold setting: The optimal threshold is determined by backtesting historical crash day data, combined with dynamic baseline adjustment.

[0015] Preferably, the model verification performed by the risk warning model construction and verification module includes: Backtesting validation: The training set and test set are divided by time, and the evaluation metrics include recall rate on crash day and accuracy rate of daily warnings; Real-time monitoring: Deploy the online learning module of the model, calculate the feature distribution drift daily, and trigger model retraining if the drift amount > 0.1; Stress testing: Simulate extreme market scenarios to verify the stability of the model under nonlinear fluctuations.

[0016] This invention provides a reference system for analyzing stock price crash risks based on stock market interaction data. Compared with existing technologies, it has the following advantages: 1. This stock price crash risk analysis reference system based on stock market interactive data identification utilizes a multimodal data acquisition and processing module. This module integrates multi-source data such as text and charts, and dynamically allocates weights using an attention mechanism, improving the comprehensiveness and accuracy of data processing. The group behavior pattern modeling and analysis module employs directed weighted graphs and the SIR model to more accurately identify key nodes and simulate the spread of negative emotions. Enhanced sentiment analysis combines multimodal and group characteristics, using an XGBoost model to output sentiment labels, enhancing the accuracy of sentiment recognition. Adversarial defense effectively filters noise interference through the generation of adversarial examples and a dual machine learning framework. The risk warning module integrates multiple factors, uses a LightGBM model to dynamically adjust weights, and ensures model stability and warning accuracy through backtesting and real-time monitoring. This application provides an analytical solution that offers inexperienced investors a reference channel for stock investment, avoiding blind following and reducing economic losses.

[0017] 2. This stock price crash risk analysis reference system, based on stock market interactive data identification, simultaneously collects multimodal data from investor interaction platforms and financial video platforms, and concurrently acquires stock market quotes and trading data, resulting in a richer and more comprehensive data source. Chart processing utilizes OCR technology to accurately extract key price levels and capital flow information; text-chart semantic matching verification ensures data consistency; emoji sentiment analysis uses a pre-trained model to accurately classify and generate a comprehensive sentiment index; voice emotion enhancement leverages an end-to-end ASR model and LSTM network to effectively predict voiceprint features and map them to sentiment intensity correction factors; data alignment and fusion synchronize data by timestamp and employ an attention mechanism to dynamically allocate weights, enabling the reasonable fusion of data from different market scenarios, improving the accuracy, comprehensiveness, and flexibility of data processing, and providing a more reliable data foundation for subsequent analysis.

[0018] 3. This stock price crash risk analysis reference system, based on stock market interaction data, uses a directed weighted graph to present user relationships in its social network construction. It comprehensively considers node attributes such as users' historical sentiment tendencies and influence scores, as well as edge weights determined by interaction frequency and sentiment consistency, to more accurately characterize user relationships. Key node identification filters out KOLs and marks sentiment mutation points, helping to capture key changes in market sentiment. Topic evolution tracking utilizes BERTopic clustering to calculate semantic drift and topic negativity index, defining topic escalation rules to promptly identify high-risk topics. Group sentiment propagation quantification constructs a SIR model to dynamically simulate the spread of negative sentiment and predict the arrival time of sentiment peaks. Overall, this module comprehensively and deeply analyzes group behavior patterns, providing strong support for accurately grasping market sentiment dynamics and providing early warnings of stock price risks, enhancing the timeliness and accuracy of the analysis.

[0019] 4. This stock price crash risk analysis reference system, based on stock market interaction data, utilizes multi-dimensional sentiment quantification combined with multimodal and group characteristics to construct vectors. XGBoost is used to output sentiment labels and confidence levels, providing a more comprehensive and accurate assessment of sentiment. In adversarial text processing, TextFooler is used to generate adversarial examples to build a test set to evaluate robustness. FGM adversarial training is employed to optimize the loss function and improve the model's resistance to interference. Counterfactual causal inference, based on the DML framework, correlates sentiment text with stock price fluctuations, controls the average processing effect of market factor estimation, and uses GANs to construct counterfactual scenarios to filter noise, making stock price crash risk warnings more accurate and reliable, and providing strong support for financial market stability.

[0020] 5. This stock price crash risk analysis reference system, based on stock market interaction data, integrates sentiment, market, and capital factors through multi-factor fusion modeling. It uses the LightGBM model to train a risk classifier, comprehensively considering various factors influencing stock price crashes and improving the accuracy of early warnings. Dynamic weight adjustment automatically adjusts factor weights according to market conditions, enabling the model to better adapt to different market environments. The early warning threshold is determined through backtesting using historical crash day data, combined with dynamic baseline adjustment, improving the sensitivity and reliability of early warnings. The model validation system includes backtesting, real-time monitoring, and stress testing, comprehensively evaluating the model's performance and ensuring its stability and accuracy under different time scales and market scenarios, providing strong support for stock price crash risk early warning. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the steps of the present invention; Figure 2 This is a system module block diagram of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] See Figures 1-2 This invention discloses a stock price crash risk analysis reference system based on stock market interaction data identification, and provides the following five technical solutions: First implementation method: includes: The multimodal data acquisition and processing module collects text, charts, emojis, and voice data from investor interactions on online platforms, and simultaneously acquires stock quotes and trading data; it extracts key price levels and fund flows from charts, performs text-chart semantic matching, conducts sentiment analysis and enhancement on emojis and voice data respectively, synchronizes data by timestamp, and dynamically allocates weights using an attention mechanism; by integrating multi-source data such as text and charts and using an attention mechanism to dynamically allocate weights, the multimodal data acquisition and processing module improves the comprehensiveness and accuracy of data processing; The group behavior pattern modeling and analysis module constructs a directed weighted graph based on user relationships, identifies key nodes and marks emotional mutation points; clusters post topics, calculates semantic drift and topic negativity index, and defines topic escalation rules; constructs an emotion propagation SIR model to simulate the speed and scope of negative emotion spread; the group behavior pattern modeling and analysis module uses directed weighted graphs and SIR models to more accurately identify key nodes and simulate the spread of negative emotions. The enhanced sentiment analysis and adversarial defense module combines multimodal data and group behavioral characteristics to construct sentiment feature vectors. It outputs sentiment classification labels and confidence scores using an XGBoost model. Adversarial examples are generated to evaluate model robustness. An adversarial training strategy is employed to optimize the loss function. Based on a dual machine learning framework, the average treatment effect is estimated, and counterfactual scenarios are constructed to filter noise interference. Enhanced sentiment analysis, combining multimodal and group characteristics, uses an XGBoost model to output sentiment labels, improving the accuracy of sentiment recognition. Adversarial defense effectively filters noise interference through the generation of adversarial examples and the dual machine learning framework. Risk warning model construction and verification module: Integrates sentiment, market, and capital factors, uses the LightGBM model to train a risk classifier, and dynamically adjusts factor weights; ensures model stability and accuracy through backtesting, real-time monitoring, and stress testing; the risk warning module integrates multiple factors, uses the LightGBM model to dynamically adjust weights, and ensures model stability and warning accuracy through backtesting and real-time monitoring.

[0024] This application does not directly predict whether the stock market will crash, nor does it combine stock market financial data analysis. It only provides a method for investors to judge the risk of stock price crash and then make targeted investments to reduce losses. By collecting interactive data from online platforms, an algorithm model is used to automatically identify interactive data, quantify and classify sentiment, and further quantify it into a risk value of stock price crash.

[0025] The second implementation method differs from the first implementation method in that the working steps of the multimodal data acquisition and processing module include: 1.1 Data source coverage: Collect text J, charts C, emoticons E, and voice V data from investor interaction platforms (stock forums, social media) and financial video platforms, and simultaneously obtain real-time market data M and historical trading data L for the corresponding stocks; In terms of data source coverage, it simultaneously collects multimodal data from investor interaction platforms and financial video platforms, and also acquires stock quotes and trading data, making the data sources richer and more comprehensive. 1.2 Multimodal Analysis Process: Chart processing: Using OCR technology, the support level (S) and resistance level (R) values ​​in the candlestick chart, and the net inflow (I) and net outflow (O) amounts in the fund flow chart, are extracted to construct a chart feature vector. I O>0 This is an indicator function that takes the value 1 when 0 > 0, and 0 otherwise. Text-Chart Semantic Matching: Perform semantic matching verification between the price point P mentioned in the text and the key price point on the chart, and calculate the Euclidean distance. If d is less than the threshold d', it is marked as consistent; otherwise, manual review is triggered. Here, k is a variable that takes a value from the set {S, R}. Emoji Sentiment Analysis: A pre-trained visual ransomware model (such as Vi-Base) is used to classify emojis, outputting emotion labels such as anger, panic, and optimism, along with probability distributions e=[e1, e2, ..., e...]. K The overall sentiment index is generated by combining the text sentiment weight w (e.g., neutral text + panic emoji = 0.3 negative weight): , where α is the emoji emotion weight (e.g., panic = 0.8) and β is the text-emoji fusion coefficient (e.g., 0.3). Speech emotion enhancement: Speech content is transcribed using an end-to-end ASR model (such as Conformer), while Mel-frequency cepstral coefficients (MFCC) features are extracted. The speech rate (s) is then predicted using an LSTM network. t Pitch p t Energy e t Voiceprint feature parameters are then mapped to emotion intensity correction factors (e.g., high pitch + rapid speech rate corresponds to panic intensity +0.2), specifically including: concatenating features into a vector: The input layer of the LSTM network is a sequence {x1, x2, ..., x}. T}, where T is the time step; the LSTM layer is used for hidden state updates: The output layer calculates the correction factor through a fully connected layer. Among them, W s This represents the weight matrix of the output layer, which maps the hidden states to correction factors; b s Here, σ is the bias term of the output layer; σ is the σ function, which compresses the output to the (0, 1) interval (if the range needs to be limited) or directly uses linear activation; the final emotion intensity is... ; Data alignment and fusion: Synchronize multimodal data by timestamp and dynamically allocate the weight of each modality using an attention mechanism (e.g., increase the weight of text when breaking negative news). , where q t For the query vector, p i v i For each modal key-value pair; To measure investor pessimism, we can follow the approach of Antweiler and Frank (2004) and use the following indicators to construct positive investor sentiment: ; Among them, Positive j，t Represents investor optimism. It is the total number of positive sentiment posts in the investor Q&A section of Company J during period t; This represents the total number of negative posts in investor inquiries for company j during period t. A higher positive value indicates a higher level of investor optimism. Secondly, consider investor pessimism and define it as the inverse of investor optimism. Specifically, taking the negative value yields: ; Invpess imism j，t The larger the value, the more pessimistic the overall investor sentiment. Chart processing utilizes OCR technology to accurately extract key price levels and capital flow information; text-chart semantic matching verification ensures data consistency; emoji sentiment analysis uses a pre-trained model to accurately classify and generate a comprehensive sentiment index; voice emotion enhancement leverages an end-to-end ASR model and LSTM network to effectively predict voiceprint features and map them to sentiment intensity correction factors; data alignment and fusion synchronize data by timestamp and use an attention mechanism to dynamically allocate weights, enabling reasonable fusion of data from different market scenarios, improving the accuracy, comprehensiveness, and flexibility of data processing, and providing a more reliable data foundation for subsequent analysis.

[0026] The third implementation method differs from the first implementation method in that the working steps of the group behavior pattern modeling and analysis module include: 2.1 Social Network Construction: A directed weighted graph is constructed based on user IDs and reply / forward relationships. Node attributes include user historical sentiment and influence score. User historical sentiment is represented by the ratio of positive to negative posts, and influence score is represented by PageRank value. Edge weights are determined by interaction frequency and sentiment consistency (e.g., if user A frequently forwards user B's negative posts, the edge weight increases). The PageRank value is calculated using the following formula: Where N represents the total number of nodes in the social network, i.e. the number of users, d is the damping coefficient, which takes a value of 0-1, and In(u) and Out(v) are the sets of incoming links for user u and outgoing links for user v, respectively. Identify key moments: Select the top 10% of users by PageRank as KOLs and mark their emotional turning points (such as a long-term neutral user suddenly posting an extremely negative post): This is used to calculate the degree of emotional change ΔE of user u at time t. u (t), and determine whether the change exceeds a statistical threshold based on the sentiment values ​​of the past 24 hours, where mean is the mean function. Indicates user u is from The sentiment value sequence up to t over the past 24 hours; std is the standard deviation function; when When this occurs, it indicates that the current emotional value at time t deviates significantly from the average emotional level of the past 24 hours, meaning that user u experienced a sudden emotional shift at time t. In this case, ΔE... u If (t) is true, it is marked as an emotional mutation point; otherwise, it is not considered that a significant emotional mutation has occurred. In building social networks, user relationships are presented using a directed weighted graph. This graph comprehensively considers node attributes such as users' historical emotional tendencies and influence scores, as well as edge weights determined by interaction frequency and emotional consistency, to more accurately depict the connections between users. Key node identification filters out KOLs and marks emotional inflection points, which helps to capture key changes in market sentiment. 2.2 Topic Evolution Tracking: BERTopic is used to cluster post topics by hourly time windows, generating topic word vectors and calculating semantic drift (e.g., the cosine similarity decreases from "performance falls short of expectations" to "financial fraud"). A sentiment lexicon is then used to calculate the topic's negative index (proportion of negative words × sentiment intensity coefficient). ;where P t Let be the set of keywords, and neg(f) be the weight of negative words (e.g., "plummeting" = 0.9). λ t As an emotion intensity attenuation factor, η=0.1, Δt is the time interval; Define topic escalation rules: When the negative index of a topic rises for two consecutive hours and the turnover rate of related stocks increases simultaneously, it is marked as a high-risk topic, triggering a backtracking of the propagation path (such as the number of levels from KOL posting to community spread). Topic evolution tracking utilizes BERTopic clustering to calculate semantic drift and topic negativity index, defines topic escalation rules, and can promptly detect high-risk topics; 2.3 Quantification of Group Emotional Spread: A Susceptible-Infected-Recovered (SIR) model for emotion spread is constructed, classifying users into three categories: uninfected (A1), affected (A2), and immune (A3). The model dynamically simulates the spread speed and coverage of negative emotions in social networks, predicting the peak time of emotions. The dynamic equation of the SIR model is as follows: ;in For the transmission rate, The recovery rate is obtained by fitting historical data (e.g.) =0.3, =0.1); The module quantifies the spread of group emotions by constructing a SIR model to dynamically simulate the diffusion of negative emotions and predict the timing of emotional peaks. Overall, this module provides a comprehensive and in-depth analysis of group behavior patterns, offering strong support for accurately grasping market sentiment dynamics and providing early warnings of stock price risks, thus enhancing the timeliness and accuracy of the analysis.

[0027] The fourth implementation method differs from the first in that the working steps of the enhanced sentiment analysis and adversarial defense module include: 3.1 Multi-dimensional Emotion Quantification: By combining multimodal data with group behavioral characteristics, an emotion feature vector is constructed: Among them, NER t The term frequency statistics vector (e.g., the number of times stock codes appear) of the text named entity recognition results; λ KOL This refers to the propagation coefficient of key nodes, which is the weight after normalization based on PageRank values; Output sentiment classification labels using the XGBoost model: ; It also outputs the confidence score, calculated using Softmax: c∈{positive, neutral, negative}; where ϕ c The original score for category c is the output of XGBoost, and j is an index variable for summation; Multi-dimensional emotion quantification combines multimodal and group features to construct vectors, and uses XGBoost to output emotion labels and confidence scores for a more comprehensive and accurate judgment of emotions; 3.2 Adversarial Text Processing: TextFooler is used to generate adversarial examples (such as replacing "plummeting" with "callback" using synonyms, and grammatical transformations from "bearish" to "bearish") to construct a test set containing implicit expressions (irony, metaphor) to evaluate the model's robustness. The adversarial training strategy is as follows: perturbation text is added during the model fine-tuning stage, adversarial noise is generated using FGM (Fast Gradient Method), and the loss function is optimized as a weighted sum of the original loss and the adversarial loss: L total =(1−b)L original +bL adv ; where L original L represents the original loss value calculated based on normal training data. adv denoted as the adversarial loss value calculated after adding perturbation text, where b is the proportion of perturbation samples; In adversarial text processing, TextFooler is used to generate adversarial examples to build a test set to evaluate robustness, and FGM adversarial training is used to optimize the loss function and improve the model's anti-interference ability. 3.3 Counterfactual Causal Reasoning: Based on the dual machine learning DML framework, sentiment text is used as the processing variable and stock price changes are used as the outcome variable. After controlling for market factors (such as the market index and industry fluctuations), the average treatment effect (ATE) is estimated. Counterfactual scenarios are constructed: the stock price path under the condition that a negative sentiment text has not been published is simulated by generative adversarial network (GAN). The difference between the actual path and the counterfactual path is compared to filter out noise interference (e.g., if the ATE confidence interval contains 0, the warning is suppressed). Counterfactual causal reasoning, based on the DML framework, associates sentiment text with stock price fluctuations, controls the average processing effect of market factor estimation, and uses GAN to construct counterfactual scenarios to filter noise, making stock price crash risk warnings more accurate and reliable, and providing strong support for financial market stability.

[0028] The fifth implementation method differs from the first implementation method in that the working steps of the risk warning model construction and verification module include: 4.1 Multi-factor fusion modeling: A feature matrix F is constructed by integrating sentiment factors, market factors, and funding factors. A risk classifier is trained using the LightGBM model. The feature matrix F is represented as follows: Among them, NegSent represents the speed of negative sentiment propagation, reflecting the extent of negative sentiment spread in the market, and is a sentiment factor; TopicNeg represents the topic negativity index, measuring the intensity of negative sentiment for a specific topic, and is a sentiment factor; VIX is the market volatility index, reflecting market uncertainty, and is a market factor; Turnover represents sudden changes in turnover rate, reflecting changes in market trading activity, and is a market factor; DiscountRate represents the discount rate for large transactions, reflecting the discount situation of large transactions and indirectly reflecting the tightness of the money supply, and is a money factor; Multi-factor fusion modeling integrates sentiment, market, and capital factors, and uses the LightGBM model to train a risk classifier, comprehensively considering multiple factors that affect stock price crashes and improving the accuracy of early warnings. Dynamic weight adjustment: Automatically adjusts factor weights based on market conditions; assuming three market conditions: bull market, bear market, and sideways market; automatically adjusts the weights of each factor B according to different market conditions. ; Dynamic weight adjustment automatically adjusts factor weights based on market conditions, enabling the model to better adapt to different market environments; 4.2 Warning threshold setting: The optimal threshold is determined by backtesting historical crash day data (e.g., triggering a warning when the model outputs a risk probability > 0.75), combined with dynamic baseline adjustment (the threshold fluctuates by ±0.05 based on recent market volatility). The optimal threshold for early warning is determined by backtesting historical crash day data and combined with dynamic baseline adjustment, which improves the sensitivity and reliability of early warning. 4.3 Model Validation System: Backtesting validation: The training set (2018-2022) and the test set (2023-2024) were divided by time. The evaluation metrics included recall rate on crash days (target >90%) and daily warning accuracy rate (target >75%). Real-time monitoring: Deploy the online learning module of the model, calculate the feature distribution drift (KS test) daily, and trigger model retraining if the drift is >0.1; Stress testing: Simulate extreme market scenarios (such as black swan events) to verify the stability of the model under nonlinear fluctuations; The model validation system includes backtesting, real-time monitoring, and stress testing, which comprehensively evaluates the model's performance and ensures its stability and accuracy under different time scales and market scenarios, providing strong support for early warning of stock price crash risks.

[0029] Furthermore, any content not described in detail in this specification is existing technology known to those skilled in the art.

[0030] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the 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.

[0031] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A stock price crash risk analysis reference system based on stock market interaction data identification, characterized in that, include: Multimodal data acquisition and processing module: Collects text, charts, emoticons, and voice data from investor interactions on the online platform, and simultaneously acquires stock quotes and transaction data; Chart processing extracts key price levels and fund flows, performs text-chart semantic matching, performs sentiment analysis and enhancement on emojis and voice respectively, synchronizes data by timestamp, and dynamically allocates weights using an attention mechanism; Group behavior pattern modeling and analysis module: Constructs a directed weighted graph based on user relationships, identifies key nodes and marks emotional mutation points; clusters post topics, calculates semantic drift and topic negativity index, defines topic escalation rules; constructs an emotion propagation SIR model to simulate the speed and scope of negative emotion spread; Enhanced sentiment analysis and adversarial defense module: Combining multimodal data and group behavior characteristics, constructing sentiment feature vectors, outputting sentiment classification labels and confidence scores through the XGBoost model; generating adversarial examples to evaluate model robustness, optimizing the loss function using an adversarial training strategy, and estimating the average treatment effect based on a dual machine learning framework, constructing counterfactual scenarios to filter noise interference; Risk warning model construction and validation module: integrates sentiment, market and capital factors, uses the LightGBM model to train a risk classifier, and dynamically adjusts the factor weights; Backtesting, real-time monitoring, and stress testing ensure the stability and accuracy of the model.

2. The stock price crash risk analysis reference system based on stock market interaction data identification according to claim 1, characterized in that: The multimodal data acquisition and processing module collects data including: text J, charts C, emoticons E, and voice V from investor interaction platforms and financial video platforms, and simultaneously obtains real-time stock quotes M and historical transaction data L for the corresponding stocks.

3. The stock price crash risk analysis reference system based on stock market interaction data identification according to claim 1, characterized in that: The multimodal data acquisition and processing module performs data processing steps including: Chart processing: Using OCR technology, the support level (S) and resistance level (R) values ​​in the candlestick chart, and the net inflow (I) and net outflow (O) amounts in the fund flow chart, are extracted to construct a chart feature vector. I O>0 This is an indicator function that takes the value 1 when 0 > 0, and 0 otherwise. Text-Chart Semantic Matching: Perform semantic matching verification between the price point P mentioned in the text and the key price point on the chart, and calculate the Euclidean distance. If d is less than the threshold d', it is marked as consistent; otherwise, manual review is triggered. Here, k is a variable that takes a value from the set {S, R}. Emoji Sentiment Analysis: A pre-trained visual ransformer model is used to classify emojis, outputting labels for anger, panic, and optimism, along with probability distributions e=[e1, e2, ..., e...]. K A comprehensive sentiment index is generated by combining the text sentiment weight w: Where α is the emotion weight of the emoji and β is the text-emoji fusion coefficient; Speech emotion enhancement: An end-to-end ASR model is used to transcribe speech content, while simultaneously extracting Mel-frequency cepstral coefficient features. An LSTM network is then used to predict the speech rate. t , pitch p t Energy e t Voiceprint feature parameters are then mapped to emotion intensity correction factors, specifically including: concatenating the features into a vector. The input layer of the LSTM network is a sequence {x1, x2, ..., x}. T }, where T is the time step; the LSTM layer is used for hidden state updates: The output layer calculates the correction factor through a fully connected layer. Among them, W s This represents the weight matrix of the output layer, which maps the hidden states to correction factors; b s Here, σ is the bias term of the output layer; σ is the σ function, which compresses the output to the (0, 1) interval or directly uses linear activation; the final emotion intensity is... ; Data alignment and fusion: Synchronize multimodal data by timestamp and dynamically allocate modality weights using an attention mechanism. , where q t For the query vector, p i v i These are the key-value pairs for each modality.

4. The stock price crash risk analysis reference system based on stock market interaction data identification according to claim 1, characterized in that: The steps of the group behavior pattern modeling and analysis module include social network construction, topic evolution tracking, and quantification of group emotion propagation. The social network construction steps include: constructing a directed weighted graph based on user ID and reply / forward relationships; node attributes include user historical sentiment tendency and influence score; user historical sentiment tendency is represented by the ratio of positive to negative posts; and influence score is represented by PageRank value; edge weights are interaction frequency and sentiment consistency; the PageRank value calculation formula is: Where N represents the total number of nodes in the social network, i.e. the number of users, d is the damping coefficient, which takes a value of 0-1, and In(u) and Out(v) are the sets of incoming links for user u and outgoing links for user v, respectively. Identify key moments: Select the top 10% of users by PageRank as KOLs and mark their emotional turning points: This is used to calculate the degree of emotional change ΔE of user u at time t. u (t), and determine whether the change exceeds a statistical threshold based on the sentiment values ​​of the past 24 hours, where mean is the mean function. Indicates user u is from The sentiment value sequence up to t over the past 24 hours; std is the standard deviation function; when When this occurs, it indicates that the current emotional value at time t deviates significantly from the average emotional level of the past 24 hours, meaning that user u experienced a sudden emotional shift at time t. In this case, ΔE... u If (t) is true, it is marked as an emotional mutation point; otherwise, it is not considered that a significant emotional mutation has occurred.

5. The stock price crash risk analysis reference system based on stock market interaction data identification according to claim 4, characterized in that: The steps for tracking topic evolution include: clustering post topics using BERTopic by hourly time windows, generating topic word vectors and calculating semantic drift, and calculating the topic negativity index using a sentiment dictionary. ;where P t Let be the set of keywords, neg(f) be the weight of negative words, and λ be the weight of negative words. t As an emotion intensity attenuation factor, η=0.1, Δt is the time interval; Define topic escalation rules: When the negative index of a topic rises for two consecutive hours and the turnover rate of related stocks increases simultaneously, it is marked as a high-risk topic, triggering a backtracking of the propagation path.

6. The stock price crash risk analysis reference system based on stock market interaction data identification according to claim 4, characterized in that: The steps for quantifying the spread of group emotions include: constructing an emotion spread SIR model, classifying users into three categories: uninvolved (A1), affected (A2), and immune (A3), dynamically simulating the spread speed and coverage of negative emotions in social networks, and predicting the arrival time of the emotion peak. The dynamic equation of the SIR model is: ;in For the transmission rate, The recovery rate is obtained by fitting historical data.

7. The stock price crash risk analysis reference system based on stock market interaction data identification according to claim 1, characterized in that: The enhanced sentiment analysis and adversarial defense module includes the following for enhanced sentiment analysis: constructing a sentiment feature vector by combining multimodal data and group behavioral characteristics. Among them, NER t λ is the word frequency statistical vector of the text named entity recognition results. KOL This refers to the propagation coefficient of key nodes, which is the weight after normalization based on PageRank values; Output sentiment classification labels using the XGBoost model: ; It also outputs the confidence score, calculated using Softmax: c∈{positive, neutral, negative}; where ϕ c Let j be the raw score of category c output by XGBoost, and j be an index variable for summation.

8. The stock price crash risk analysis reference system based on stock market interaction data identification according to claim 7, characterized in that: The enhanced sentiment analysis and adversarial defense module includes the following adversarial defense features: Adversarial text processing: TextFooler is used to generate adversarial examples, constructing a test set containing implicit expressions to evaluate model robustness; the adversarial training strategy is as follows: perturbation text is added during the model fine-tuning stage, adversarial noise is generated using FGM, and the loss function is optimized as a weighted sum of the original loss and the adversarial loss: L total =(1−b)L original +bL adv ; where L original L represents the original loss value calculated based on normal training data. adv denoted as the adversarial loss value calculated after adding perturbation text, where b is the proportion of perturbation samples; Counterfactual causal reasoning: Based on the dual machine learning DML framework, sentiment text is used as the processing variable and stock price change is used as the outcome variable. After controlling for market factors, the average processing effect is estimated. Counterfactual scenarios are constructed: the stock price path under the condition that a negative sentiment text has not been published is simulated by generative adversarial network GAN. The difference between the actual path and the counterfactual path is compared to filter out noise interference.

9. A stock price crash risk analysis reference system based on stock market interaction data identification according to claim 1, characterized in that: The risk warning model construction and verification module includes the following steps: integrating sentiment factors, market factors, and capital factors to construct a feature matrix F, and training a risk classifier using the LightGBM model; the feature matrix F is represented as follows: Among them, NegSent represents the speed of negative sentiment spread, reflecting the extent of negative sentiment diffusion in the market, and belongs to the sentiment factor; TopicNeg represents the topic negativity index, measuring the intensity of negative sentiment for a specific topic, and belongs to the sentiment factor; VIX is the market volatility index, reflecting market uncertainty, and belongs to the market factor; Turnover represents sudden changes in turnover rate, reflecting changes in market trading activity, and belongs to the market factor; DiscountRate represents the discount rate for large transactions, reflecting the discount situation of large transactions, indirectly reflecting the tightness of the funding situation, and belongs to the funding factor; Dynamic weight adjustment: Automatically adjusts factor weights based on market conditions; assuming three market conditions: bull market, bear market, and sideways market; automatically adjusts the weights of each factor B according to different market conditions. ； Warning threshold setting: The optimal threshold is determined by backtesting historical crash day data, combined with dynamic baseline adjustment.

10. A stock price crash risk analysis reference system based on stock market interaction data identification according to claim 9, characterized in that: The risk warning model construction and verification module performs model verification, including the following: Backtesting validation: The training set and test set are divided by time, and the evaluation metrics include recall rate on crash day and accuracy rate of daily warnings; Real-time monitoring: Deploy the online learning module of the model, calculate the feature distribution drift daily, and trigger model retraining if the drift amount > 0.1; Stress testing: Simulate extreme market scenarios to verify the stability of the model under nonlinear fluctuations.

Images