Stock time sequence event-based knowledge graph construction method and device, storage medium and computer device

Abstract

The application discloses a stock time sequence event-based knowledge graph construction method and device, a storage medium and computer equipment. The method comprises the following steps: collecting stock time sequence events corresponding to a target stock, matching historical reference events in a historical prototype library by using a large model to determine an initial influence period, combining the industry attribute, volatility and historical similar event influence period of the target stock to predict a correction period and adaptively correct the initial influence period, obtaining a predicted influence period, and then determining the triggering time, invalidation time and event intensity of the event, finally constructing a stock node and an event node in a knowledge graph, establishing a correlation through an edge with a time sequence attribute label, and forming a structured dynamic event influence knowledge graph.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a method, apparatus, storage medium, and computer equipment for constructing a knowledge graph based on stock time-series events. Background Technology

[0002] In the fields of financial investment and quantitative analysis, the stock market is continuously affected by various time-series events (such as financial report releases, policy changes, executive shareholding changes, and industry-wide emergencies). Investors and researchers rely heavily on accurate judgments of the cycle and intensity of these events to support trading decisions, risk management, and portfolio optimization. With the explosive growth of market data, constructing knowledge graphs that dynamically reflect the correlation between events and stock prices has become a key technological direction for improving the efficiency of financial information processing and the level of intelligence in investment analysis.

[0003] Currently, the construction of knowledge graphs for stock events mainly relies on manual sorting and expert annotation. Specifically, analysts manually extract event elements by reading announcements, news, and research reports, and set the start and end times and weights of the event's impact based on subjective experience, before associating and storing event nodes with stock nodes. This manual construction method is inefficient and struggles to handle massive, high-frequency event flows. Summary of the Invention

[0004] In view of this, embodiments of this application provide a method, apparatus, storage medium, and computer device for constructing a knowledge graph based on stock time-series events.

[0005] According to one aspect of this application, a method for constructing a knowledge graph based on stock time-series events is provided, the method comprising: Collect stock time-series events corresponding to the target stock; The large model obtains historical reference events that match the stock time series event from the historical prototype library, and determines the initial impact period corresponding to the stock time series event based on the event impact period corresponding to the historical reference event. The industry attributes, stock volatility, impact cycles of historical similar events corresponding to the stock time series event, and event types of the target stock are obtained. Based on the industry attributes, stock volatility, and impact cycles of historical similar events, the correction cycle of the stock time series event is predicted, and the initial impact cycle is corrected according to the correction cycle to obtain the predicted impact cycle. Determine the trigger time of the stock time series event, calculate the failure time based on the trigger time and the predicted impact period, and determine the event intensity based on the predicted impact period; In the knowledge graph, the stock node corresponding to the target stock is determined, the time-series event node corresponding to the stock time-series event is constructed, and the edge between the stock node and the time-series event node is constructed. The edge is labeled based on the generation time, the trigger time, the expiration time and the event intensity of the stock time-series event.

[0006] According to another aspect of this application, a knowledge graph construction apparatus based on stock time-series events is provided, the apparatus comprising: The event acquisition module is used to collect stock time-series events corresponding to the target stock; The impact cycle calculation module is used to obtain historical reference events matching the stock time series event from the historical prototype library through a large model, and determine the initial impact cycle corresponding to the stock time series event based on the event impact cycle corresponding to the historical reference events; obtain the industry attributes, stock volatility, impact cycles of similar historical events corresponding to the stock time series event, and event type of the target stock; predict the correction cycle of the stock time series event based on the industry attributes, stock volatility, and impact cycles of similar historical events; and correct the initial impact cycle based on the correction cycle to obtain the predicted impact cycle. The event intensity calculation module is used to determine the trigger time of the stock time series event, calculate the failure time based on the trigger time and the predicted impact period, and determine the event intensity based on the predicted impact period. The knowledge graph construction module is used to determine the stock node corresponding to the target stock in the knowledge graph, construct the time-series event node corresponding to the stock time-series event, and construct the edge between the stock node and the time-series event node, and label the edge based on the generation time, trigger time, failure time and event intensity of the stock time-series event.

[0007] According to another aspect of this application, a storage medium is provided that stores a computer program thereon, which, when executed by a processor, implements the above-described knowledge graph construction method based on stock time-series events.

[0008] According to another aspect of this application, a computer device is provided, including a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, wherein the processor executes the program to implement the above-described knowledge graph construction method based on stock time-series events.

[0009] By employing the above technical solutions, this application provides a method, apparatus, storage medium, and computer device for constructing a knowledge graph based on stock time-series events. This method collects stock time-series events corresponding to a target stock, uses a large model to match historical reference events in a historical prototype library to determine the initial impact cycle, and then combines the target stock's industry attributes, volatility, and the impact cycles of similar historical events to predict and correct the cycle, adaptively correcting the initial impact cycle to obtain the predicted impact cycle. This allows for the determination of the event's trigger time, expiration time, and event intensity. Finally, stock nodes and event nodes are constructed in the knowledge graph and linked through edges labeled with time-series attributes, forming a structured, dynamic event impact knowledge graph. This method automates the entire process of constructing a stock time-series event knowledge graph, effectively replacing traditional manual labeling methods and improving construction efficiency. By introducing large-model semantic matching and individual stock characteristic correction mechanisms, the prediction of event impact cycles becomes more accurate and adaptive. Simultaneously, the quantitative labeling of event intensity and expiration time imbues the graph with rich time-series semantic information, providing a high-quality data foundation for subsequent stock association analysis, event transmission path mining, and investment decision support.

[0010] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0011] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 The illustration shows a flowchart of a knowledge graph construction method based on stock time-series events provided in an embodiment of this application; Figure 2 This illustration shows a flowchart of another knowledge graph construction method based on stock time-series events provided in an embodiment of this application; Figure 3 This illustration shows a flowchart of another knowledge graph construction method based on stock time-series events provided in an embodiment of this application; Figure 4 This illustration shows a schematic diagram of a knowledge graph construction device based on stock time-series events provided in an embodiment of this application. Figure 5 A schematic diagram of the device structure of a computer device provided in an embodiment of this application is shown. Detailed Implementation

[0012] The present application will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the embodiments of the present application can be combined with each other.

[0013] This embodiment provides a method for constructing a knowledge graph based on stock time-series events, such as... Figure 1 As shown, the method includes: Step 101: Collect the stock time series events corresponding to the target stock.

[0014] Among them, stock time-series events refer to stock-related event information with a clear occurrence time on the timeline, which includes event type attributes (such as financial report release, performance forecast, executive increase or decrease in shareholding, signing of major contracts, regulatory inquiries, etc.) and event occurrence time attributes; target stocks refer to specific stocks to be constructed into the knowledge graph, which are uniquely identified by stock code or stock name.

[0015] Specifically, the target stock for which the knowledge graph is to be built is first determined. Based on the stock code or name, the corresponding stock time-series events are collected from multi-source data interfaces. Data sources include, but are not limited to, listed company announcement interfaces, financial news push interfaces, exchange disclosure data interfaces, and third-party financial data service interfaces. In one implementation, a timed polling method is used to actively retrieve new event information from various data sources at preset time intervals (e.g., every minute or hour). In another implementation, a message subscription method is used, accessing the real-time push service provided by the data source to receive event data instantly when an event occurs.

[0016] In one optional embodiment, collecting stock time-series events corresponding to the target stock includes: collecting stock time-series events of different event types corresponding to the target stock in multiple markets, performing cross-market time alignment on the stock time-series events, and unifying the generation time and trigger time of the stock time-series events to a preset standard time; wherein, the market of the stock time-series event is marked in the time-series event node.

[0017] In this embodiment, "multiple markets" refers to stock trading markets in different countries or regions, including but not limited to the A-share market, the Hong Kong stock market, and the US stock market, each with different time zone attributes and trading rules. Cross-market time alignment refers to uniformly converting event time information from different markets, carrying different time zone timestamps, into a preset standard time (such as Coordinated Universal Time UTC or Beijing Time) to eliminate the interference caused by time zone differences in the construction of time sequence relationships. The preset standard time refers to a unified time benchmark pre-set by the system, used to standardize the storage and calculation of the time attributes of all events. Specifically, the target stock to be constructed for the knowledge graph is first determined. Based on the stock code or stock name of the target stock, the corresponding stock time sequence events are collected from multi-source data interfaces. The data sources collected include, but are not limited to, listed company announcement interfaces, financial news push interfaces, exchange disclosure data interfaces, and third-party financial data service interfaces. For example, for a target stock listed on both the A-share and Hong Kong stock markets, events occurring in each market are collected from their respective announcement systems. After the data collection is completed, cross-market time alignment is performed on stock time series events from different markets: the original timestamp of each event record and the time zone information of its market are identified, and the original timestamp is converted into a preset standard time according to the time zone conversion rules (e.g., uniformly converted to UTC time or Beijing time), and the converted standard time is used as the unified storage format for the generation time and trigger time of the event.

[0018] Step 102: Obtain historical reference events that match the stock time series event from the historical prototype library using the large model, and determine the initial impact period corresponding to the stock time series event based on the event impact period corresponding to the historical reference event.

[0019] The historical prototype library refers to a pre-built dataset that stores records of historical stock events and their impact cycles. This library contains a large number of historical reference event samples with labeled impact cycles. The initial impact cycle refers to the estimated duration of the event's impact, determined based on the impact cycles of historical reference events, without adjustment for individual stock characteristics.

[0020] Specifically, the event description text of stock time-series events is input into the large model. The large model semantically encodes the event description, generating a corresponding semantic feature vector. Then, it performs vector retrieval in a historical prototype database, calculating the similarity between this vector and the semantic vectors of each historical reference event in the database. The top-K historical similar events are extracted as historical reference events, and an initial impact period is generated based on the average impact period of the historical reference events. .

[0021] Step 103: Obtain the industry attributes, stock volatility, impact cycle and event type of the historical similar events corresponding to the stock time series event of the target stock; predict the correction cycle of the stock time series event based on the industry attributes, stock volatility and impact cycle of the historical similar events; and correct the initial impact cycle based on the correction cycle to obtain the predicted impact cycle.

[0022] Among these, industry attribute refers to the identifier information of the industry category to which the target stock belongs, reflecting the overall operating rules and event response characteristics of the industry in which the stock operates. Stock volatility is a measure of the degree of price fluctuation of the target stock, reflecting the price sensitivity and market activity of the stock. The impact cycle of similar historical events refers to the statistical value of the impact cycle of historical events of the same event type as the current stock time series event in the historical prototype library. Event type refers to the category label to which the stock time series event belongs, used to classify the event and index the corresponding historical similar event data. The correction cycle is the impact cycle calculated based on the individual characteristics of the target stock (industry attribute, volatility) and the performance of similar historical events. The predicted impact cycle is the final estimated impact cycle value obtained after correcting the initial impact cycle for individual stock characteristics.

[0023] Specifically, the process begins by acquiring the target stock's industry attributes, stock volatility, and the event type corresponding to the current stock's time-series events. Industry attributes can be obtained by querying a stock basic information database; stock volatility is calculated based on the target stock's historical price series, with one implementation using the standard deviation of historical daily returns as the volatility measure. Then, based on the event type, all similar events are retrieved from a historical prototype database, and the impact cycle records for each of these historical similar events are extracted. Their statistical values ​​(such as median, mode, or mean) are calculated as the impact cycle of the historical similar events. In one implementation, only historical similar events belonging to the same industry as the current stock are selected for statistical analysis to enhance comparability. In another implementation, all historical similar events are selected for statistical analysis, and the statistical values ​​are weighted and adjusted based on industry similarity. After obtaining the industry attributes, stock volatility, and the impact cycle of historical similar events, a revised cycle is predicted based on these three pieces of information. For example, a pre-trained cycle prediction model (such as a regression model or decision tree model) is used, with the industry attribute code, stock volatility value, and the impact cycle of historical similar events as input features. The model outputs the revised cycle. After obtaining the revised cycle, the initial impact cycle is revised based on this revised cycle. In one implementation, the initial impact period and the correction period are weighted and summed to obtain the predicted impact period. Through the individual stock characteristic correction mechanism and multi-strategy prediction design in this step, the predicted impact period can adaptively reflect the different response characteristics of different stocks to the same event.

[0024] In one optional embodiment, the initial impact period is corrected according to the correction period to obtain the predicted impact period, including: obtaining the adaptive weight coefficient corresponding to the event type of the stock time series event; and performing a weighted summation of the initial impact period and the correction period according to the adaptive weight coefficient to obtain the predicted impact period.

[0025] In this embodiment, the adaptive weighting coefficient refers to an adjustable parameter, preset or dynamically determined for different event types, used to adjust the proportion of the initial impact period and the correction period in the weighted summation. This coefficient reflects the differences in the dependence of different event types on historical reference experience and their sensitivity to individual stock characteristics. For example, for events with strong market universality (such as the release of macroeconomic data), the initial impact period (based on historical reference events) has higher reference value, and the adaptive weighting coefficient assigns a larger weight to the initial impact period; for events with strong individual stock specificity (such as changes in corporate governance), the adjustment role of the correction period (based on individual stock characteristics) is more critical, and the adaptive weighting coefficient is adjusted accordingly to highlight the contribution of the correction period. After determining the adaptive weights, the predicted impact period is calculated using the following formula: , To predict the impact cycle, For adaptive weighting coefficients, For the initial prediction period, This refers to the prediction period for this type of event.

[0026] Step 104: Determine the trigger time of the stock time series event, calculate the failure time based on the trigger time and the predicted impact period, and determine the event intensity based on the predicted impact period.

[0027] The trigger time refers to the point at which a stock-series event officially begins to have a market impact. The expiration time refers to the point at which the impact of the stock-series event ends, calculated by superimposing the trigger time on the predicted impact period. The event intensity is a measure of the degree to which a stock-series event affects its related stocks, determined based on the length of the predicted impact period.

[0028] Specifically, the trigger time of the stock-related time-series event is first determined. For example, the trigger delay is determined based on the event type. For instance, for financial report events, the actual disclosure date of the financial report is the trigger time, not the announcement date; therefore, the specific disclosure date needs to be parsed from the event description as the trigger time. For example, calibration is performed based on market trading hours. If the event occurs outside of trading hours, the trigger time is postponed to the opening time of the next trading session. After determining the trigger time, the expiration time is calculated based on the trigger time and the predicted impact period. In one implementation, the expiration time is obtained by adding the duration corresponding to the predicted impact period to the trigger time. In another implementation, the period calculation method is determined based on the calendar type (calendar day or trading day). If trading days are used, only trading days are included in the period count. Then, the event intensity is determined based on the predicted impact period.

[0029] In an optional embodiment, determining the event intensity based on the predicted impact period includes: determining the impact duration type of the stock time series event based on its event type, wherein the impact duration type is a long-term impact type or a short-term impact type; obtaining the impact type weight corresponding to the impact duration type; and determining the event intensity based on the predicted impact period and the impact type weight using an event intensity formula, wherein the event intensity formula is expressed as follows: , Indicates the intensity of the event. Indicates the influence type weight. This indicates the predicted impact period.

[0030] In this embodiment, the impact duration type refers to the classification of the duration of the impact of stock time-series events based on their event type, divided into two categories: long-term impact type and short-term impact type. Events corresponding to the long-term impact type typically trigger structural changes at the fundamental or industry trend level (such as major asset restructuring, industry policy shifts, etc.), and their impact lasts for a relatively long time. Events corresponding to the short-term impact type typically trigger market sentiment or short-term supply and demand fluctuations (such as short-term earnings forecast revisions, block trades, etc.), and their impact lasts for a relatively short time. The impact type weight refers to the intensity adjustment coefficient preset according to different impact duration types, used to assign differentiated basic intensity labels to events of different duration types, reflecting the difference in the impact strength of long-term and short-term events on stock prices within a unit impact period.

[0031] First, the event type of the current stock time series event is obtained, and its corresponding impact period type is determined based on this event type. For example, a mapping table between event types and impact period types is pre-constructed. For instance, event types such as mergers and acquisitions and policy / regulatory changes are mapped to long-term impact types, while events such as preliminary earnings reports and executive statements are mapped to short-term impact types. This is determined directly by looking up the table. After determining the impact period type, the corresponding impact type weight is obtained. In one implementation, long-term impact types correspond to larger impact type weight values, and short-term impact types correspond to smaller impact type weight values. The relative size between the two is set empirically; for example, short-term types have a weight of 1.0, and long-term types have a weight of 0.6. Then, the event intensity is determined using an event intensity formula based on the predicted impact period and the impact type weight. This formula is: ,in Indicates the intensity of the event. Indicates the influence type weight. This formula represents the predicted impact period. By taking the natural logarithm of the predicted impact period, adding one, and then multiplying the reciprocal by the impact type weight, it establishes a non-linear negative correlation between event intensity and the predicted impact period. That is, with a fixed impact type weight, the longer the predicted impact period, the lower the event intensity per unit time, consistent with the actual law that event impact diminishes marginally over time. Through the determination of impact duration type and the design of the logarithmic decay formula, the determination of event intensity not only distinguishes the fundamental strength differences between long-term and short-term events but also objectively reflects the natural law of event influence decreasing over time through the non-linear decay relationship.

[0032] Step 105: Determine the stock node corresponding to the target stock in the knowledge graph, construct the time-series event node corresponding to the stock time-series event, and construct the edge between the stock node and the time-series event node. Label the edge based on the generation time, trigger time, failure time and event intensity of the stock time-series event.

[0033] In this context, a stock node refers to a graph node in the knowledge graph that represents the target stock entity, storing its basic attribute information. A time-series event node refers to a graph node in the knowledge graph that represents the stock's time-series event entity, storing the event's content attribute information. Generation time refers to the original timestamp of the stock's time-series event in the data source. Trigger time refers to the point in time when the event begins to have a market impact. Expiration time refers to the point in time when the event's impact ends. Event intensity refers to a measure of the degree of the event's impact.

[0034] Specifically, the process begins by locating or creating a stock node corresponding to the target stock within the knowledge graph. If the stock node already exists in the knowledge graph, it is retrieved directly; otherwise, a new stock node is created and populated with the stock's basic attribute information (such as stock code, stock name, industry category, etc.). Next, the time-series event node corresponding to the current stock's time-series event is constructed. For example, each stock time-series event is independently represented by a time-series event node, storing attribute information such as event type, event description summary, and event occurrence time. Then, the edges between the stock node and the time-series event node are constructed. After the edges are built, they are labeled based on the occurrence time, trigger time, expiration time, and event intensity of the stock time-series event.

[0035] By applying the technical solution of this embodiment, stock time-series events corresponding to the target stock are collected. A large model is used to match historical reference events in a historical prototype library to determine the initial impact cycle. Then, the industry attributes, volatility, and impact cycles of similar historical events of the target stock are combined to predict and correct the cycle, adaptively adjusting the initial impact cycle to obtain the predicted impact cycle. This allows for the determination of the event's trigger time, expiration time, and event intensity. Finally, stock nodes and event nodes are constructed in a knowledge graph and linked through edges labeled with time-series attributes, forming a structured dynamic event impact knowledge graph. This method automates the entire process of constructing a stock time-series event knowledge graph, effectively replacing traditional manual labeling and improving construction efficiency. By introducing large-model semantic matching and individual stock characteristic correction mechanisms, the prediction of event impact cycles becomes more accurate and adaptive. Simultaneously, the quantitative labeling of event intensity and expiration time imbues the graph with rich time-series semantic information, providing a high-quality data foundation for subsequent stock correlation analysis, event transmission path mining, and investment decision support.

[0036] In an optional embodiment, when the current time reaches a preset update time, based on the expiration time of the edges in the knowledge graph, edges with expiration times greater than or equal to the preset update time, along with the nodes connected by these edges, are filtered out to form the knowledge graph. The event intensity corresponding to the edges in the filtered knowledge graph is then corrected using an event intensity correction formula, wherein the event intensity correction formula is expressed as follows: , This indicates the corrected event intensity, and T represents the current time. Indicates the trigger time.

[0037] In this embodiment, the preset update time refers to the knowledge graph maintenance trigger time pre-set by the system. This can be a fixed time point (e.g., after the market closes daily, before the market opens on Mondays) or a fixed time interval (e.g., every 6 hours, every 24 hours). Filtering refers to filtering edges in the knowledge graph based on their expiration time attributes, retaining only edges that meet preset conditions and their connected nodes. Specifically, when the current time reaches the preset update time, the update process is initiated. First, based on the expiration time attributes of the edges in the knowledge graph, all edges in the graph are traversed, filtering out edges whose expiration time is greater than or equal to the current preset update time, as well as the stock nodes and time-series event nodes connected to these edges, forming a filtered subset of the knowledge graph. After filtering, the event intensity corresponding to each edge in the retained knowledge graph is corrected. An event intensity correction formula is used. Perform a correction operation, where the current time T is the current system time, and the trigger time is... Read the attribute values ​​already labeled in the edges to predict the influence cycle. Similarly, it is read from the edge attributes. Through periodic filtering and exponential decay correction design, the event intensity information in the knowledge graph can be automatically updated over time, accurately reflecting the dynamic characteristics of the event's impact gradually diminishing as the market operates.

[0038] Furthermore, as a refinement and extension of the specific implementation of the above embodiments, and to fully illustrate the specific implementation process of this embodiment, another method for constructing a knowledge graph based on stock time-series events is provided, such as... Figure 2 As shown, the method includes: Step 201: Obtain the daily abnormal return rate of the target stock after the start of the trigger time; Step 202: When the daily abnormal return rate is lower than a preset threshold for a preset number of consecutive days, determine that the event impact of the stock time series event has terminated, record the actual failure time of the stock time series event, and determine the actual impact period of the stock time series event based on the actual failure time and the trigger time. Step 203: Compare the actual impact period and the predicted impact period of the stock time series events; Step 204: If the deviation between the predicted impact period and the actual impact period is greater than the preset deviation, then the failure time and event intensity of the edge corresponding to the stock time series event are recalculated and labeled.

[0039] In this embodiment, the daily abnormal return of the target stock after the trigger time is first obtained. In one implementation, the daily abnormal return is calculated using a market model method, that is, based on the regression relationship between the historical return of the target stock before the event and the return of the market index, the expected return for each trading day after the event is estimated, and the daily abnormal return is obtained by subtracting the expected return from the actual return. Then, when the daily abnormal return is lower than a preset threshold for a preset number of consecutive days, the event impact of the stock's time-series event is determined to have terminated, and this moment is recorded as the actual failure time. In one implementation, the preset number of days is set to a fixed value (e.g., three consecutive trading days), and the preset threshold is set to a fixed percentage (e.g., the daily abnormal return is lower than -0.5% for a consecutive period), both of which are pre-configured based on business experience. After the termination condition is triggered, the actual failure time is recorded, and the actual impact period is calculated based on the actual failure time and the trigger time, that is, the actual failure time minus the duration corresponding to the trigger time. After obtaining the actual impact period, the actual impact period is compared with the predicted impact period. For example, the absolute difference or relative difference between the two is calculated as a deviation measure. If the deviation between the predicted impact period and the actual impact period is greater than a preset deviation threshold, a correction process is triggered. For example, a preset deviation threshold is set in absolute days (e.g., a deviation exceeding 3 trading days triggers a correction). Once the correction condition is triggered, the failure time and event intensity of the edges corresponding to the stock time-series events are recalculated and labeled. For example, the actual failure time is directly replaced with the original failure time in the edge, and the event intensity is recalculated based on this actual failure time, ensuring that the edge labeling matches the actual observation results. Through the daily abnormal return tracking and prediction result verification design of this embodiment, the event impact labeling in the knowledge graph possesses closed-loop feedback correction capabilities. The prediction results are verified and corrected using actual market data, thus improving the accuracy of the current graph labeling.

[0040] In an optional embodiment, predicting the correction period of the stock time series event based on the industry attribute, the stock volatility, and the impact period of similar historical events includes: using the industry attribute, stock volatility, the impact period of similar historical events, and the event type as input features, performing regression prediction using a pre-trained period prediction model, and outputting the correction period of the stock time series event; wherein, the stock volatility is determined based on the standard deviation of the target stock's historical daily return series within the most recent preset time period; After recalculating and labeling the failure time and event intensity of the edges corresponding to the stock time series event, the method further includes: constructing training samples based on the industry attribute, the stock volatility, the impact cycle of the historical similar events, the event type, and the actual impact cycle of the stock time series event, and iteratively training the cycle prediction model; and storing the stock time series event and its actual impact cycle as new records in the historical prototype library.

[0041] In this embodiment, the cycle prediction model refers to a pre-trained machine learning model used to predict and correct the cycle based on input features. It achieves intelligent estimation of the current event's correction cycle by learning the mapping relationship between industry attributes, volatility, the impact cycle of similar historical events, event type, and actual correction amount in historical events. The historical daily return sequence refers to a data sequence composed of the returns of the target stock's closing price relative to the previous trading day's closing price within the most recent preset time period, arranged in chronological order. The standard deviation is a measure of the dispersion calculated from the above return sequence; a larger value indicates more volatile stock prices and a higher sensitivity to events. Training samples refer to data units used for iterative training of the cycle prediction model, containing input features (industry attributes, stock volatility, the impact cycle of similar historical events, and event type) and corresponding label values ​​(actual impact cycle).

[0042] Specifically, after recalculating and labeling the failure time and event intensity of the edges corresponding to the stock time series event, the following model iteration and knowledge accumulation operations can be performed. On the one hand, training samples are constructed using the actual observation data of the event to iteratively train the cycle prediction model. Specifically, industry attributes, stock volatility, the impact cycle of similar historical events, and event type are used as input features, and the actual impact cycle obtained by tracking daily abnormal returns in this verification process is used as the label value to form a complete training sample. In one implementation, the training sample is added to the model's historical training set, triggering the incremental training process of the cycle prediction model. The model fine-tunes its parameters based on the new samples on the basis of the original parameters, without having to start full training from scratch. In another implementation, the new samples are merged with the historical training samples and the model is fully retrained to ensure that the overall model fits optimally. For example, when the number of accumulated training samples reaches a preset batch threshold (such as adding 100 samples), a unified iterative training is triggered to avoid the consumption of computational resources caused by frequent training. On the other hand, the stock time series event and its actual impact cycle are stored as new records in the historical prototype library. For example, the complete information of the current event (event type, event description, industry attributes, actual impact period, etc.) can be directly added to the historical prototype library as a new historical reference event for matching with subsequent similar events. Through iterative training of the periodic prediction model and dynamic updating of the historical prototype library, the ability to correct periodic predictions has the characteristic of continuous self-evolution. The actual observation results of each event can be fed back to the prediction system, driving the optimization of model parameters and the enrichment of the reference experience library.

[0043] In one alternative embodiment, such as Figure 3 As shown, the method further includes: Step 301: Receive user query statement, parse the user query statement, and extract the entity target, market scope, and time window corresponding to the user query statement; based on the entity target and the market scope, filter target event nodes and target stock nodes in the knowledge graph, and determine knowledge graph segments based on the target event nodes and the target stock nodes.

[0044] In this context, a user query refers to a query request entered by a user in natural language or structured form to retrieve information related to stock events within the knowledge graph. The entity target refers to the identifier of the stock entity to be queried, extracted from the user query; this can be the stock code or name of a single or multiple stocks. The market scope refers to the market domain defined in the query, such as the A-share market, Hong Kong stock market, US stock market, or cross-market. The time window refers to the time range defined in the query, used to constrain the time interval of the retrieved events. A knowledge graph fragment refers to a subgraph selected from the complete knowledge graph based on the query conditions; it contains target event nodes, target stock nodes, and edges between nodes related to the query.

[0045] Specifically, the system first receives a query from the user. For example, the query might be input in natural language, such as "What events affected a certain stock in the past month and their spillover effects on other related stocks?" After receiving the query, it is parsed to extract query elements. In one implementation, natural language processing tools are used to segment the query, identify entities, and extract relationships, identifying key information such as stock names, market names, and timeframes. In another implementation, a large model is used to perform semantic understanding on the query, directly extracting entity targets, market scope, and time windows from the statement. Based on the extracted entity targets and market scope, the corresponding target stock node is located in the knowledge graph. Next, target event nodes are filtered based on market scope and time windows. In one implementation, the target stock node is first located based on the entity target, then all event nodes connected to that stock node through edges are obtained, and a second filtering is performed based on market scope and time windows to narrow the search scope and improve efficiency. After obtaining the target stock nodes and target event nodes, knowledge graph segments are determined based on these nodes and the edge relationships between them. For example, all target stock nodes, target event nodes, and the edges between them are extracted to form a subgraph as a knowledge graph fragment.

[0046] Step 302: Based on the query time corresponding to the user query statement, adjust the event intensity corresponding to each edge in the knowledge graph segment.

[0047] The query time refers to the current time when the user initiates the query. It is used to correct the timeliness of events in the knowledge graph fragments so that the query results reflect the impact status of events up to the current moment.

[0048] Specifically, the query time corresponding to the user's query is obtained. After obtaining the query time, the event intensity corresponding to each edge in the knowledge graph segment is adjusted. For example, the aforementioned exponential decay formula is used. The correction is made, where the current time T is replaced with the query time and the trigger time. and predict the impact cycle The corrected event intensity is calculated by reading the annotation attributes from the edge.

[0049] Step 303: Determine the target stock node pairs in the knowledge graph segment connected by an intermediary event node, and obtain the corrected event intensity of the edge between each target stock node and the corresponding target event node in the target stock node pair.

[0050] In this context, a mediator event node refers to an event node in a knowledge graph segment that simultaneously connects at least two target stock nodes, acting as a mediator bridge in the graph structure to establish relationships between stocks. A target stock node pair refers to a combination of two different target stock nodes that are linked through the same mediator event node.

[0051] Specifically, the process involves identifying target stock node pairs connected by an intermediary event node within a knowledge graph fragment. For example, iterating through all event nodes in the knowledge graph fragment, for each event node, obtaining a list of all connected target stock nodes. When this list contains two or more target stock nodes, these nodes are paired to form target stock node pairs. After identifying the target stock node pairs, the modified event strength of the edge between each target stock node and its corresponding target event node in the pair is obtained.

[0052] Step 304: Calculate the comprehensive strength of the target stock node pair using the comprehensive strength calculation formula, wherein the comprehensive strength calculation formula is expressed as follows: , This represents the combined strength of target stock node i and target stock node j. Indicates the number of intermediary event nodes. This represents the adjusted event strength between the target stock node i and the k-th intermediary event node. This represents the adjusted event strength between the target stock node j and the k-th intermediary event node. This represents the event type weight corresponding to the k-th intermediary event node. This indicates cross-market weighting.

[0053] The overall strength refers to the measure of the overall correlation between two target stock nodes connected by intermediary event nodes. It integrates the contributions of multiple intermediary event nodes as well as the moderating effects of event type and cross-market factors. Event type weight refers to a pre-set weighting coefficient based on the event type to which the intermediary event node belongs; different types of events have different importance in transmitting the correlation between stocks. Cross-market weight refers to a weighting coefficient used to adjust the strength of cross-market event correlation, reflecting the differences in the event transmission effect between different markets.

[0054] Specifically, the overall strength of the target stock node pair is calculated using a comprehensive strength calculation formula. This formula represents... ,in, This represents the combined strength of target stock node i and target stock node j. Indicates the number of intermediary event nodes. , Let represent the adjusted event strengths of target stock nodes i and j and the kth intermediary event node, respectively. This represents the event type weight corresponding to the k-th intermediary event node. This indicates cross-market weighting. Through a multi-factor comprehensive strength calculation formula, the correlation strength between stocks is comprehensively and objectively quantified, reflecting the combined influence of factors such as event intensity, event type importance, and cross-market attenuation.

[0055] Step 305: Using a large model, based on the knowledge graph fragment, the corrected event intensity, the target stock node pair, and the comprehensive intensity of the target stock node pair, generate query results, wherein the query results include event description, event propagation path, and the comprehensive intensity of the target stock node pair.

[0056] The query results refer to the final returned information generated based on the user's query statement, which includes an event description, the event propagation path, and the overall strength of the target stock node pairs. The event propagation path refers to the graph structure path connecting the target stock nodes through intermediary event nodes, visually demonstrating how the event has a correlation and impact on one or more stocks.

[0057] Specifically, first, the large model generates event descriptions based on knowledge graph fragments. In one implementation, the large model extracts attribute information (such as event type, event description summary, and occurrence time) of intermediary event nodes from the knowledge graph fragments and organizes it into a fluent natural language description explaining which events occurred. In another implementation, the event description also includes an overview of the specific direction and degree of the event's impact on the stock. Then, the large model generates event propagation paths based on the node and edge relationships in the knowledge graph fragments. For example, the propagation chain "Stock A - Event X - Stock B" is described in text form, explaining that Stock A and Stock B are associated through Event X. Finally, the large model integrates the overall strength of the target stock node pairs into the query results. For example, the overall strength is presented directly in numerical form in the results, supplemented by natural language descriptions (such as "The overall association strength between the two stocks is X"). The large model integrates the event descriptions, event propagation paths, and the overall strength of the target stock node pairs into a structured query result output. In one implementation, the query results are presented in formatted text form. This allows structured knowledge graph data to be transformed into natural language query results that are easy for users to understand and use, improving information readability and user experience.

[0058] Furthermore, as Figure 1 In terms of specific implementation, this application provides a knowledge graph construction device based on stock time-series events, such as... Figure 4 As shown, the device includes: The event acquisition module is used to collect stock time-series events corresponding to the target stock; The impact cycle calculation module is used to obtain historical reference events matching the stock time series event from the historical prototype library through a large model, and determine the initial impact cycle corresponding to the stock time series event based on the event impact cycle corresponding to the historical reference events; obtain the industry attributes, stock volatility, impact cycles of similar historical events corresponding to the stock time series event, and event type of the target stock; predict the correction cycle of the stock time series event based on the industry attributes, stock volatility, and impact cycles of similar historical events; and correct the initial impact cycle based on the correction cycle to obtain the predicted impact cycle. The event intensity calculation module is used to determine the trigger time of the stock time series event, calculate the failure time based on the trigger time and the predicted impact period, and determine the event intensity based on the predicted impact period. The knowledge graph construction module is used to determine the stock node corresponding to the target stock in the knowledge graph, construct the time-series event node corresponding to the stock time-series event, and construct the edge between the stock node and the time-series event node, and label the edge based on the generation time, trigger time, failure time and event intensity of the stock time-series event.

[0059] In an optional embodiment, the event acquisition module is configured to: Collect stock time-series events of different event types corresponding to the target stock in multiple markets, perform cross-market time alignment on the stock time-series events, and unify the generation time and trigger time of the stock time-series events to a preset standard time; wherein, the market of the stock time-series event is marked in the time-series event node.

[0060] In an optional embodiment, the influence cycle calculation module is used for: Obtain the adaptive weight coefficients corresponding to the event types of the stock time-series events; The predicted influence period is obtained by weighting and summing the initial influence period and the correction period according to the adaptive weighting coefficients.

[0061] In an optional embodiment, the event intensity calculation module is configured to: Based on the event type of the stock time series event, determine the impact period type to which the stock time series event belongs, wherein the impact period type is a long-term impact type or a short-term impact type; Obtain the impact type weight corresponding to the impact period type; The event intensity is determined using an event intensity formula based on the predicted impact period and the impact type weight, wherein the event intensity formula is expressed as follows: , Indicates the intensity of the event. Indicates the influence type weight. This indicates the predicted impact period.

[0062] In an optional embodiment, the map construction module is further configured to: When the current time reaches the preset update time, based on the expiration time of the edges in the knowledge graph, the knowledge graph consisting of edges whose expiration time is greater than or equal to the preset update time and the nodes connected by the edges is filtered out. The event intensity of edges in the filtered knowledge graph is corrected using an event intensity correction formula, wherein the event intensity correction formula is expressed as follows: , This indicates the corrected event intensity, and T represents the current time. Indicates the trigger time.

[0063] In an optional embodiment, the map construction module is further configured to: Obtain the daily abnormal return of the target stock after the start of the trigger time; When the daily abnormal return rate is lower than a preset threshold for a preset number of consecutive days, the event impact of the stock time series event is determined to have terminated. The actual failure time of the stock time series event is recorded, and the actual impact period of the stock time series event is determined based on the actual failure time and the trigger time. Compare the actual impact period and the predicted impact period of the aforementioned stock time-series events; If the deviation between the predicted impact period and the actual impact period is greater than the preset deviation, the failure time and event intensity of the edge corresponding to the stock time series event will be recalculated and labeled.

[0064] In an optional embodiment, the influence cycle calculation module is used for: The industry attributes, stock volatility, the impact cycle of similar historical events, and the event type are used as input features. A pre-trained cycle prediction model is used to perform regression prediction, and the corrected cycle of the stock time series event is output. The stock volatility is determined based on the standard deviation of the historical daily return series of the target stock within the most recent preset time period. The device further includes: a feedback module, used for: Training samples are constructed based on the industry attributes, stock volatility, the impact cycle of similar historical events, the event type, and the actual impact cycle of the stock time series event. The cycle prediction model is then iteratively trained. Additionally, the stock time series event and its actual impact cycle are stored as new records in the historical prototype library.

[0065] In an optional embodiment, the apparatus further includes: a query module, configured to: The system receives user query statements, parses the user query statements, and extracts the entity target, market scope, and time window corresponding to the user query statements; based on the entity target and the market scope, it filters target event nodes and target stock nodes in the knowledge graph, and determines knowledge graph segments based on the target event nodes and stock nodes. Based on the query time corresponding to the user query statement, the event intensity corresponding to each edge in the knowledge graph segment is adjusted; Identify the target stock node pairs in the knowledge graph segment that are connected by an intermediary event node, and obtain the corrected event intensity of the edge between each target stock node and the corresponding target event node in the target stock node pair; The comprehensive strength of the target stock node pair is calculated using the comprehensive strength calculation formula, wherein the comprehensive strength calculation formula is expressed as follows: , This represents the combined strength of target stock node i and target stock node j. Indicates the number of intermediary event nodes. This represents the adjusted event strength between the target stock node i and the k-th intermediary event node. This represents the adjusted event strength between the target stock node j and the k-th intermediary event node. This represents the event type weight corresponding to the k-th intermediary event node. Indicates cross-market weighting; Using a large model, based on the knowledge graph fragment, the corrected event intensity, the target stock node pair, and the comprehensive intensity of the target stock node pair, query results are generated, wherein the query results include event description, event propagation path, and the comprehensive intensity of the target stock node pair.

[0066] It should be noted that other corresponding descriptions of the functional units involved in the knowledge graph construction device based on stock time-series events provided in this application embodiment can be found in the following references. Figures 1 to 3 The corresponding descriptions in the method will not be repeated here.

[0067] This application also provides a computer device, which may specifically be a personal computer, a server, a network device, etc. Figure 5 As shown, the computer device includes a bus, a processor, memory, and a communication interface, and may also include an input / output interface and a display device. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores location information. The network interface allows communication with external terminals via a network connection. When the computer program is executed by the processor, it implements the steps in the various method embodiments.

[0068] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0069] In one embodiment, a computer-readable storage medium is provided, which may be non-volatile or volatile, having stored thereon a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0070] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0071] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, graphics processors, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0072] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0073] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for constructing a knowledge graph based on stock time-series events, characterized in that, The method includes: Collect stock time-series events corresponding to the target stock; The large model obtains historical reference events that match the stock time series event from the historical prototype library, and determines the initial impact period corresponding to the stock time series event based on the event impact period corresponding to the historical reference event. The industry attributes, stock volatility, impact cycles of historical similar events corresponding to the stock time series event, and event types of the target stock are obtained. Based on the industry attributes, stock volatility, and impact cycles of historical similar events, the correction cycle of the stock time series event is predicted, and the initial impact cycle is corrected according to the correction cycle to obtain the predicted impact cycle. Determine the trigger time of the stock time series event, calculate the failure time based on the trigger time and the predicted impact period, and determine the event intensity based on the predicted impact period; In the knowledge graph, the stock node corresponding to the target stock is determined, the time-series event node corresponding to the stock time-series event is constructed, and the edge between the stock node and the time-series event node is constructed. The edge is labeled based on the generation time, the trigger time, the expiration time and the event intensity of the stock time-series event.

2. The method according to claim 1, characterized in that, Collect stock time-series events corresponding to the target stock, including: Collect stock time-series events of different event types corresponding to the target stock in multiple markets, perform cross-market time alignment on the stock time-series events, and unify the generation time and trigger time of the stock time-series events to a preset standard time; wherein, the market of the stock time-series event is marked in the time-series event node.

3. The method according to claim 2, characterized in that, The initial impact period is corrected based on the correction period to obtain the predicted impact period, including: Obtain the adaptive weight coefficients corresponding to the event types of the stock time-series events; The predicted influence period is obtained by weighting and summing the initial influence period and the correction period according to the adaptive weighting coefficients.

4. The method according to claim 2, characterized in that, Determining the event intensity based on the predicted impact period includes: Based on the event type of the stock time series event, determine the impact period type to which the stock time series event belongs, wherein the impact period type is a long-term impact type or a short-term impact type; Obtain the impact type weight corresponding to the impact period type; The event intensity is determined using an event intensity formula based on the predicted impact period and the impact type weight, wherein the event intensity formula is expressed as follows: , Indicates the intensity of the event. Indicates the influence type weight. This indicates the predicted impact period.

5. The method according to claim 4, characterized in that, After labeling the edges based on the generation time, trigger time, expiration time, and event intensity of the stock time-series event, the method further includes: When the current time reaches the preset update time, based on the expiration time of the edges in the knowledge graph, the knowledge graph consisting of edges whose expiration time is greater than or equal to the preset update time and the nodes connected by the edges is filtered out. The event intensity of edges in the filtered knowledge graph is corrected using an event intensity correction formula, wherein the event intensity correction formula is expressed as follows: , This indicates the corrected event intensity, and T represents the current time. Indicates the trigger time.

6. The method according to claim 1, characterized in that, The method further includes: Obtain the daily abnormal return of the target stock after the start of the trigger time; When the daily abnormal return rate is lower than a preset threshold for a preset number of consecutive days, the event impact of the stock time series event is determined to have terminated. The actual failure time of the stock time series event is recorded, and the actual impact period of the stock time series event is determined based on the actual failure time and the trigger time. Compare the actual impact period and the predicted impact period of the aforementioned stock time-series events; If the deviation between the predicted impact period and the actual impact period is greater than the preset deviation, the failure time and event intensity of the edge corresponding to the stock time series event will be recalculated and labeled.

7. The method according to claim 6, characterized in that, Based on the industry attributes, stock volatility, and the impact cycles of similar historical events, the correction cycle of the stock time series event is predicted, including: The industry attributes, stock volatility, the impact cycle of similar historical events, and the event type are used as input features. A pre-trained cycle prediction model is used to perform regression prediction, and the corrected cycle of the stock time series event is output. The stock volatility is determined based on the standard deviation of the historical daily return series of the target stock within the most recent preset time period. After recalculating and labeling the failure time and event intensity of the edges corresponding to the stock time series events, the method further includes: Training samples are constructed based on the industry attributes, stock volatility, the impact cycle of similar historical events, the event type, and the actual impact cycle of the stock time series event. The cycle prediction model is then iteratively trained. Additionally, the stock time series event and its actual impact cycle are stored as new records in the historical prototype library.

8. The method according to claim 2, characterized in that, The method further includes: The system receives user query statements, parses the user query statements, and extracts the entity target, market scope, and time window corresponding to the user query statements; based on the entity target and the market scope, it filters target event nodes and target stock nodes in the knowledge graph, and determines knowledge graph segments based on the target event nodes and stock nodes. Based on the query time corresponding to the user query statement, the event intensity corresponding to each edge in the knowledge graph segment is adjusted; Identify the target stock node pairs in the knowledge graph segment that are connected by an intermediary event node, and obtain the corrected event intensity of the edge between each target stock node and the corresponding target event node in the target stock node pair; The comprehensive strength of the target stock node pair is calculated using the comprehensive strength calculation formula, wherein the comprehensive strength calculation formula is expressed as follows: , This represents the combined strength of target stock node i and target stock node j. Indicates the number of intermediary event nodes. This represents the adjusted event strength between the target stock node i and the k-th intermediary event node. This represents the adjusted event strength between the target stock node j and the k-th intermediary event node. This represents the event type weight corresponding to the k-th intermediary event node. Indicates cross-market weighting; Using a large model, based on the knowledge graph fragment, the corrected event intensity, the target stock node pair, and the comprehensive intensity of the target stock node pair, query results are generated, wherein the query results include event description, event propagation path, and the comprehensive intensity of the target stock node pair.

9. A knowledge graph construction device based on stock time-series events, characterized in that, The device includes: The event acquisition module is used to collect stock time-series events corresponding to the target stock; The impact cycle calculation module is used to obtain historical reference events matching the stock time series event from the historical prototype library through a large model, and determine the initial impact cycle corresponding to the stock time series event based on the event impact cycle corresponding to the historical reference events; obtain the industry attributes, stock volatility, impact cycles of similar historical events corresponding to the stock time series event, and event type of the target stock; predict the correction cycle of the stock time series event based on the industry attributes, stock volatility, and impact cycles of similar historical events; and correct the initial impact cycle based on the correction cycle to obtain the predicted impact cycle. The event intensity calculation module is used to determine the trigger time of the stock time series event, calculate the failure time based on the trigger time and the predicted impact period, and determine the event intensity based on the predicted impact period. The knowledge graph construction module is used to determine the stock node corresponding to the target stock in the knowledge graph, construct the time-series event node corresponding to the stock time-series event, and construct the edge between the stock node and the time-series event node, and label the edge based on the generation time, trigger time, failure time and event intensity of the stock time-series event.

10. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 8.

11. A computer device, comprising a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 8.

Images