Internet quantitative investment consultation generation method and system based on real-time market linkage

Abstract

The application provides an internet quantitative investment consultation generation method and system based on real-time market linkage, relates to the field of financial investment, and solves the technical problems that in the prior art, real-time market, quantitative signals and user risk characteristics are not effectively linked, resulting in investment consultation generation lag, poor adaptability and difficulty in verification. The method comprises the following steps: constructing a real-time market data pool and a user investment context vector; using a multi-scale sliding window to calculate a market dynamic index vector and inputting the market dynamic index vector into a pre-trained adaptive event recognition model to identify and generate market event data; generating quantitative signal data according to the corresponding financial target of the market event data; generating consultation task data through a causal inference model and generating investment consultation text; and outputting the investment consultation text that passes the verification to a user terminal. The application can improve the response speed, personalized adaptation degree and verifiability of the investment consultation content to real-time market.

Description

Technical Field

[0001] This application relates to the field of financial investment, and in particular to a method and system for generating quantitative investment information based on real-time market data. Background Technology

[0002] With the development of internet finance platforms and robo-advisory services, users can view real-time market data for financial instruments and receive investment advice through various terminals. Current technologies typically push financial information to users based on user profiles, reading preferences, or portfolio data, or use quantitative systems to calculate technical indicators and generate strategy signals based on real-time market data. However, existing solutions suffer from a lack of stable linkage between real-time market data and the advice generation process. This results in the inability to generate timely advice corresponding to market events such as abnormal fluctuations, trend breakouts, or industry-wide shifts. Furthermore, the lack of adaptation between quantitative signals and the user's context can easily lead to inconsistencies between content and user needs. Additionally, the intelligently generated financial advice lacks traceability and verifiability mechanisms, potentially resulting in inaccurate market data, insufficient risk warnings, or overly definitive statements.

[0003] Therefore, it is necessary to propose a method and system for generating internet-based quantitative investment advice based on real-time market data. Summary of the Invention

[0004] This application provides a method and system for generating quantitative investment advice based on real-time market data linkage, which solves the technical problems in the prior art where the lack of effective linkage between real-time market data, quantitative signals and user risk characteristics leads to delayed investment advice generation, poor adaptability and difficulty in verification.

[0005] To achieve the above objectives, this application adopts the following technical solution: Firstly, it provides a method for generating internet-based quantitative investment advice based on real-time market data, including: Collect real-time market data from multiple sources and build a real-time market data pool; Acquire basic and behavioral data of target users to generate user investment context vectors; the basic data includes user attribute information, including invested financial products, holding data, self-selected targets data, and risk tolerance level; the behavioral data includes user operation records on the platform. Based on the real-time market data pool, dynamic market indicator vectors at different time granularities are calculated and input into a pre-trained adaptive event recognition model to identify and generate market event data. Based on the financial instruments corresponding to the market event data, quantitative signal data is generated by combining historical market data, quantitative factor library and strategy rules; Based on the user investment context vector, the market event data, and the quantitative signal data, the potential causal effects of the user's investment portfolio are assessed, and consulting task data is generated. Based on the aforementioned consulting task data, a consulting generation template and a language generation model are invoked to generate investment consulting text; The investment advice text is verified, and the verified investment advice text is output to the user terminal.

[0006] In conjunction with the first aspect above, in one possible implementation, the step of acquiring the target user's basic data and behavioral data to generate a user investment context vector includes: Based on the behavioral data, a dynamic intent map representing users' short-term investment intentions and their correlations is constructed through natural language processing and graph neural networks; Based on the attribute information and real-time market data, a real-time risk status indicator representing the user's current actual risk exposure level is calculated. The basic data, the dynamic intent graph, and the real-time risk status indicators are used as multimodal inputs and processed through a cross-modal attention fusion network to generate a user investment context vector that integrates the user's long-term profile, short-term intent, and real-time risk status.

[0007] In conjunction with the first aspect above, in one possible implementation, the method for obtaining the risk status indicator includes: Using the tradable financial products in the attribute information as nodes, and using the real-time market data to calculate the dynamic correlation between any two nodes as edge weights, a holding-watching coupled risk network is constructed. For each node in the coupled risk network, an implicit volatility surface is fitted to its corresponding real-time market data, and a multi-dimensional risk feature vector is extracted from it. Based on the topology of the coupled risk network, the multidimensional risk feature vectors of each node are weighted and aggregated to obtain the overall risk features of the user combination. The overall risk characteristics are compared with the preset risk preference vector of the risk tolerance level to generate a real-time risk status index that includes absolute risk level and relative risk fit.

[0008] In conjunction with the first aspect above, in one possible implementation, the method for analyzing the market dynamic indicator vector includes: Construct a window pool containing multiple heterogeneous sliding windows with different time granularities; wherein the time length of each window is not uniformly distributed and contains at least one asymmetric window; For the real-time market data stream of the target financial product, a set of market microstructure features are calculated in parallel within each sliding window. The market microstructure features include at least two of the following: order flow imbalance, bid-ask depth change rate, and tick-by-tick directional ratio. Real-time monitoring of global market status parameters, including overall market volatility and liquidity indicators; Based on the global market state parameters, the current weight coefficient of each sliding window in the window pool is dynamically calculated through a preset adaptive weight function; Based on the current weighting coefficients, the microstructural features of the market data output from each sliding window are weighted and fused to generate a multidimensional dynamic indicator vector representing the current market dynamics.

[0009] In conjunction with the first aspect above, in one possible implementation, the pre-trained adaptive event recognition model includes: The fused market dynamic indicator vectors generated by continuous time steps are organized into a market dynamic indicator sequence in chronological order. Using multi-source heterogeneous external data, a multi-dimensional soft label is synthesized for each time segment in the market dynamic indicator sequence. The multi-dimensional soft label includes event type, event intensity confidence, and event duration estimation. A dual-channel neural network model is constructed; wherein the dual channels include a backbone channel and a market state perception channel; the backbone channel is used to process the dynamic market indicator sequence to extract time-series features, and the market state perception channel is used to receive global market state parameters and generate dynamic gating signals; Based on the multidimensional soft labels and the composite loss function, the dual-channel neural network model is trained end-to-end to obtain a pre-trained adaptive event recognition model; the composite loss function includes event type classification loss, event intensity regression loss, and event duration prediction loss.

[0010] In conjunction with the first aspect above, in one possible implementation, the method for generating the quantized signal data includes: Based on the financial target, retrieve historical market data within a preset time window from the historical market database, and retrieve a set of quantitative factors that match the event type from the quantitative factor library; Based on a pre-configured strategy rule engine, the historical market data, the quantitative factors, and the market event data are fused and analyzed in multiple dimensions; wherein, the strategy rule engine includes condition-action mapping rules, which are used to determine the trading logic based on the event type and factor combination; Based on the results of the multidimensional fusion analysis, structured quantized signal data is generated.

[0011] In conjunction with the first aspect above, in one possible implementation, the method for obtaining the consultation task data includes: A causal inference model is constructed, which uses the user's investment context vector as a condition variable and models the market event data and quantitative signal data as intervention variables respectively. The difference in the distribution of user portfolio returns with and without the intervention is calculated by counterfactual reasoning. The potential causal effect of the market events and quantitative signals on the user portfolio is evaluated. The potential causal effect includes the impact on portfolio volatility, maximum drawdown, expected return and position correlation. Based on the assessment results of the potential causal effects and combined with the preset consultation triggering rules, structured consultation task data is automatically generated.

[0012] In conjunction with the first aspect above, in one possible implementation, the verification of the investment advisory text includes: Perform multi-dimensional automatic verification on the investment consultation text, the multi-dimensional automatic verification including: Compliance verification: Based on a pre-built knowledge base of financial regulatory rules, detect whether there are any illegal terms, profit promises or misleading statements in the text; Consistency verification: Compare the suggested content in the investment advice text with the causal effect summary and recommended actions in the original advice task data to verify logical consistency; Data timeliness verification: Check whether the interval between the market event timestamp and quantitative signal generation time cited in the investment consultation text and the current time is within a preset valid window; Risk Adaptability Verification: Based on the risk preference level in the user's investment context vector, determine whether the operational risk level suggested in the investment advice text exceeds the user's acceptable range; When all multi-dimensional automatic verifications pass, the investment consultation text will be encrypted and pushed to the user's terminal. If any verification item fails, the investment advisory text shall be corrected according to the failed verification item, and the corrected investment advisory text shall be re-verified.

[0013] In conjunction with the first aspect above, in one possible implementation, the construction of the real-time market data pool includes: Subscribe to or poll to collect raw real-time market data from multiple market data sources for at least one financial instrument; The raw real-time market data is mapped and aligned with timestamps to form preliminary standardized data; Outlier filtering is performed on the preliminary standardized data, and cross-validation is performed on the preliminary standardized data from different market data sources that are for the same financial instrument and the same market data time to identify conflicting data. Based on the historical performance data of each market data source, a dynamic evaluation model for market data source quality is constructed. The historical performance data includes the historical delay distribution and historical price deviation distribution of each market data source under different preset market conditions. The dynamic evaluation model for market data source quality is configured to output the dynamic credibility weight of each market data source in real time according to the current market condition. Using the dynamic credibility weight, the conflicting data is adaptively weighted and fused to generate fused market data; The fused market data, the preliminary standardized data that has passed cross-validation and is conflict-free, and the data marked as abnormal or pending review are all stored in the real-time market data pool, and each piece of data in the pool is associated with its source information and fusion identifier.

[0014] Secondly, this application provides an internet-based quantitative investment consulting generation system based on real-time market data linkage, comprising: an acquisition module, an identification module, a quantification module, a consulting module, and a verification module; wherein, the acquisition module is used to collect multi-source real-time market data to construct a real-time market data pool; acquire the target user's basic data and behavioral data, and generate a user investment context vector; the identification module is used to calculate dynamic market indicator vectors at different time granularities based on the real-time market data pool, and input them into a pre-trained adaptive event recognition model to identify and generate market event data; the quantification module is used to generate quantitative signal data based on the financial targets corresponding to the market event data, combined with historical market data, a quantitative factor library, and strategy rules; the consulting module is used to evaluate the potential causal effects of the user's investment portfolio based on the user investment context vector, the market event data, and the quantitative signal data, and generate consulting task data; and based on the consulting task data, call a consulting generation template and a language generation model to generate investment consulting text; the verification module is used to verify the investment consulting text and output the verified investment consulting text to the user terminal.

[0015] This application constructs a real-time market data pool and introduces a multi-scale sliding window and an adaptive event recognition model to achieve automatic perception and triggering of market events, significantly improving the responsiveness of investment consulting content to market changes. Based on this, the system deeply integrates the identified market events with a quantitative factor library and a strategy rule library to generate structured quantitative signals containing signal direction, strength, risk level, confidence level, and basis. It then uses a causal inference model to assess the causal impact on users' investment portfolios, ensuring that the generated consulting content not only reflects market phenomena but also possesses professional and interpretable quantitative logic support. Simultaneously, the system combines user investment context vectors to perform dual risk and semantic adaptation of the consulting content, achieving personalized output for users with different risk preferences, portfolio structures, and focus areas. To ensure content reliability, the system also designs a multi-dimensional automatic verification mechanism covering compliance, consistency, timeliness, and risk adaptability, effectively reducing risks such as factual errors, missing evidence, or overly certain statements. Furthermore, by binding traceable data identifiers to key text segments, the system enables end-to-end tracking from the final consultation content back to the original market events, quantitative signals, user context, and generated templates, significantly improving the auditability, verifiability, and iterability of the generated content. Attached Figure Description

[0016] Figure 1 A flowchart illustrating an internet-based quantitative investment advisory generation method based on real-time market data linkage, provided as an embodiment of this application; Figure 2 A flowchart illustrating a user investment context vector generation method provided in this application embodiment; Figure 3 This is a flowchart illustrating another training method for an adaptive event recognition model provided in an embodiment of this application. Detailed Implementation

[0017] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. 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.

[0018] Example 1, as Figure 1 As shown, this embodiment provides a method for generating internet-based quantitative investment information based on real-time market data. This method aims to address the technical problems in existing technologies, such as the disconnect between real-time market changes and the generation of investment information content, insufficient adaptation of quantitative signals to user risk characteristics, and a lack of verifiable evidence for the generated text. The method includes the following steps: S201. Collect real-time market data from multiple sources, perform standardized processing, and construct a real-time market data pool.

[0019] Specifically, real-time market data comes from multiple sources, including stock exchanges, financial data service providers, and internet finance platforms. The data content may include, but is not limited to, the latest price, price change, trading volume, turnover, order book bid and ask prices and volume, and individual transaction data of financial instruments such as stocks, funds, bonds, futures, options, foreign exchange, and indices.

[0020] The standardization process includes mapping fields from data sources to unify them into a preset data format; aligning timestamps to unify time information from different sources to the same time base; and filtering outliers to remove obviously erroneous or invalid data. Based on historical performance data from each market data source, a dynamic evaluation model for market data source quality is constructed. Historical performance data includes the historical latency distribution and historical price deviation distribution of each market data source under different preset market conditions. The dynamic evaluation model is configured to output dynamic credibility weights for each market data source in real time based on the current market condition. Using these dynamic credibility weights, conflicting data is adaptively weighted and fused to generate fused market data. The fused market data, the preliminary standardized data that has passed cross-validation and is conflict-free, and the data marked as abnormal or pending review are all stored in a real-time market data pool, with each data point in the pool associated with its source information and fusion identifier.

[0021] S202. Obtain the target user's basic data and behavioral data, and generate the user's investment context vector.

[0022] Specifically, this step aims to build a personalized user profile. The foundational data consists of static user attributes, such as age range, investment experience, risk tolerance level, portfolio holdings, and selected investment targets. Behavioral data comprises dynamic records of user activity on the platform, such as historical consultation interaction data (e.g., past questions asked, feedback on consultation content), and platform behavior data (e.g., duration of browsing market data pages, selected investment targets, and search keywords). By comprehensively analyzing this data, a user investment context vector can be generated, representing the user's long-term investment preferences, short-term focus, and current risk profile. This data serves as a crucial input for subsequent risk adaptation and personalized content generation.

[0023] S203. Based on the real-time market data pool, a multi-scale sliding window is used to calculate the dynamic index vectors of market data at different time granularities. The dynamic index vectors of market data at different time granularities are then input into a pre-trained adaptive event recognition model to identify and generate market event data.

[0024] Among them, market event data is structured data that represents specific structural behaviors or abnormal states in the financial market.

[0025] Specifically, a multi-scale sliding window refers to using multiple windows of different time lengths (e.g., 1 minute, 5 minutes, 15 minutes, 1 hour, etc.) to slide on the real-time market data stream. Within each window, a set of market dynamic indicators are calculated, such as price change rate, volume amplification factor, volatility, turnover rate, etc., thereby forming a vector of market dynamic indicators at different time granularities.

[0026] S204. Based on the financial instruments corresponding to the market event data, combined with historical market data, quantitative factor library and strategy rules, generate quantitative signal data.

[0027] Specifically, upon identifying a market event, the system will lock onto its associated financial instrument and retrieve historical data for that instrument within a preset time window from the historical market database. Simultaneously, it will search the quantitative factor library for a set of quantitative factors matching the event type, such as technical factors, funding factors, and fundamental factors. Based on a pre-configured strategy rule engine, including rules for trend following, mean reversion, and breakout identification, the system will perform multi-dimensional fusion analysis of historical market data, quantitative factors, and current market event data. The analysis results will be structured into quantitative signal data, which may include signal direction, signal strength, risk level, confidence level, and the basis for signal generation.

[0028] S205. Based on user investment context vectors, market event data, and quantitative signal data, evaluate the potential causal effects of market events and quantitative signals on user portfolios through a causal inference model, and generate consulting task data based on the potential causal effects.

[0029] Among them, the consultation task data consists of structured instruction information that provides users with personalized investment advice or risk warnings.

[0030] S206. Based on the consulting task data, call the consulting generation template and language generation model to generate investment consulting text.

[0031] The consultation document includes suggestions for the user's investment portfolio, related financial instruments, and operational guidelines.

[0032] S207. Verify the investment consultation text. When the verification is successful, output the investment consultation text to the user terminal.

[0033] Based on the above technical solutions, this application proposes an efficient and intelligent method and system for generating quantitative investment advice. By constructing a real-time market data pool and combining a multi-scale sliding window with an adaptive event recognition model, the system automatically captures market changes and triggers advice generation, significantly improving response speed. The system links identified market events with a quantitative factor library and a strategy rule library to generate structured quantitative signals containing signal direction, strength, risk level, confidence level, and their basis. It then uses a causal inference model to assess the impact on the user's investment portfolio, ensuring the advice content is both professionally in-depth and logically interpretable. Simultaneously, the system integrates contextual information such as the user's dynamic intent graph and real-time risk status to personalize the risk level and expression granularity of the generated content, meeting the diverse needs of different investors. To ensure output quality, the system also introduces a multi-dimensional automatic verification mechanism covering compliance, consistency, data timeliness, and risk matching, effectively preventing issues such as factual errors, missing evidence, or excessive expression. In addition, key text fragments are all associated with unique data identifiers, supporting the tracing back from the final consultation content to the original market events, quantitative signals, user status, and generated templates, thereby enhancing the verifiability and auditability of the content and the system's continuous optimization capabilities.

[0034] Example 2, based on Example 1, such as Figure 2 As shown, this embodiment elaborates on the specific implementation of S202, namely, obtaining the target user's basic data and behavioral data to generate a user investment context vector. This step aims to construct a context vector that can comprehensively and dynamically represent the user's investment status, providing accurate input for subsequent personalized consultation tasks, and includes the following sub-steps: S301, based on behavioral data, constructs a dynamic intent graph that represents users' short-term investment intentions and their related relationships through natural language processing and graph neural networks.

[0035] Specifically, behavioral data includes users' historical consultation interaction data and platform behavior data, such as past user question texts, clicks and dwell time on consultation content, search keywords, and sequences of market data pages viewed. Natural language processing (NLP) techniques are used to process this textual behavioral data, extracting key entities (such as financial instrument names, industry terms, and investment strategy keywords) and sentiment tendencies. These entities are used as nodes, and the user's behavior sequences of continuously engaging with these entities within a single session or short period are used as edges to construct a dynamic graph structure. Graph neural networks are used to learn the embedding representations of the nodes in this graph structure. These embedding vectors not only encode the attributes of the entities themselves but also the co-occurrence and temporal relationships between entities in the user's behavior sequence.

[0036] S302. Based on position data, selected stock data, risk tolerance level, and real-time market data, calculate a real-time risk status indicator that represents the user's current actual risk exposure level.

[0037] The methods for obtaining risk status indicators further include: S3021. Construct a portfolio-focused coupled risk network. Specifically, tradable financial products such as stocks and funds from the user's portfolio data, as well as financial products that the user focuses on from their watchlist data, are used as nodes in the network. Using the historical return series of these financial products from a real-time market data pool, the dynamic correlation between any two nodes is calculated (e.g., using a rolling window Pearson correlation coefficient or a Dynamic Conditional Correlation Coefficient (DCC) model), and this correlation value is used as the weight of the edge connecting the two nodes. This network not only includes assets directly held by the user but also assets that the user focuses on but does not hold, thus more comprehensively reflecting the user's risk exposure and potential risk contagion paths.

[0038] S3022. For each node in the coupled risk network, fit the implied volatility surface using its corresponding real-time market data, and extract a multi-dimensional risk feature vector from it.

[0039] Implied volatility surfaces are an important tool in the options market for reflecting the expected future volatility of an underlying asset. For underlying assets such as stocks, they can be fitted using their corresponding option chain data; for underlying assets without direct option data, surrogate models can be constructed based on historical volatility, industry volatility indices, and other data for estimation. From the fitted implied volatility surface, multiple dimensions of risk characteristics can be extracted, such as the current at-the-money implied volatility level, volatility skewness (reflecting the market's expectation of tail risk), and term structure slope (reflecting the market's expectation of future volatility changes). These characteristics together constitute the multidimensional risk feature vector of that node.

[0040] S3023. Based on the topology of the coupled risk network, the multidimensional risk feature vectors of each node are weighted and aggregated to obtain the overall risk features of the user portfolio.

[0041] The aggregation process takes into account the network topology; for example, a graph attention mechanism can be used to dynamically allocate aggregation weights based on the degree centrality of nodes (the number of connecting edges) or the weight of edges (the strength of correlation). This means that assets highly correlated with the user's holdings or watchlist, and with significant risk characteristics, will have a greater impact on the overall risk profile. Through this weighted aggregation, the resulting overall risk profile can reflect the comprehensive risk exposure of the user's entire investment portfolio under complex relationships.

[0042] S3024. Compare the overall risk characteristics with the preset risk preference vector of the risk tolerance level to generate a real-time risk status indicator that includes the absolute risk level and the relative risk fit.

[0043] Among them, risk tolerance level is a static attribute determined by the user during account opening or risk assessment. It can be mapped to a preset risk preference vector, which defines the acceptable thresholds for the user in dimensions such as volatility, maximum drawdown, and downside risk. The calculated overall risk characteristic vector is compared with the risk preference vector dimension by dimension; for example, cosine similarity or Euclidean distance is calculated to generate a comprehensive real-time risk status indicator.

[0044] S3025. The basic data, dynamic intent graph and real-time risk status indicators are used as multimodal inputs and processed through a cross-modal attention fusion network to generate a user investment context vector that integrates the user's long-term profile, short-term intent and real-time risk status.

[0045] The basic data can be encoded as static feature vectors. The dynamic intent graph outputs a graph-level embedding vector through a graph neural network, representing the user's short-term intent. The real-time risk status indicator is itself a multi-dimensional vector. These three data modalities (static attributes, graph structure, and numerical vectors) are integrated through a cross-modal attention fusion network. The core mechanism of this network is attention, which allows the model to dynamically focus on the information most relevant to the current task in different modalities when generating the final context vector. For example, when a high-volatility market event is identified, the attention mechanism may focus more on the parts of the real-time risk status indicator and the dynamic intent graph that are relevant to high-risk targets, while relatively weakening the general preferences in the long-term profile. Through this adaptive fusion, the generated user investment context vector can accurately reflect the user's overall investment status in a specific market context, providing high-quality conditional variables for subsequent causal inference models to evaluate the impact of events on the user's portfolio.

[0046] Based on the above technical solution, this embodiment establishes a multi-dimensional and dynamic user investment context representation mechanism. This mechanism not only considers users' static attributes and long-term preferences, but more importantly, it captures real-time changes in user intent through dynamic intent mapping. Furthermore, by coupling risk networks and implied volatility surface analysis, it accurately quantifies the true risk exposure of user portfolios in the current market. This in-depth user profiling lays a solid foundation for generating highly personalized, risk-adaptive investment advisory content.

[0047] Example 3, based on Example 1, elaborates in detail the specific analysis method for calculating the dynamic market indicator vector based on the real-time market data pool in step S203. This method aims to establish a multi-scale, adaptive dynamic market indicator generation mechanism, including the following steps: S401. Construct a window pool containing multiple heterogeneous sliding windows with different time granularities.

[0048] The time lengths of each window are not uniformly distributed and include at least one asymmetric window.

[0049] Specifically, a heterogeneous sliding window refers to a window pool where the sliding windows are non-uniformly distributed in terms of time length and contain at least one asymmetric window. For example, a preferred window pool might contain five windows with time lengths of 1 minute, 5 minutes, 15 minutes, 1 hour, and 4 hours. The 1-minute and 5-minute windows are short-time granularities used to capture high-frequency trading behavior and instantaneous price shocks; the 15-minute and 1-hour windows are medium-time granularities used to identify intraday trends and swings; and the 4-hour window is a long-time granularity used to observe broader market sentiment and fund flows. An asymmetric window means that the start and end times of the window are not strictly aligned to the hour or fixed intervals. For example, a window can be set to look back 37 minutes from the current moment to avoid coinciding with fixed market opening and closing times, thereby reducing the interference of periodic noise on indicator calculations. This heterogeneous window pool design, compared to using a single time window or multiple uniformly distributed windows, can simultaneously capture dynamic market information from different time scales, avoiding the risk of missing key market patterns due to inappropriate window selection.

[0050] S402. For the real-time market data stream of the target financial product, calculate a set of market microstructure features in parallel within each sliding window.

[0051] Among them, market microstructure characteristics refer to indicators that can reflect deeper information such as market order flow, liquidity, and price formation mechanisms.

[0052] In this embodiment, the set of features includes, but is not limited to, at least two of the following: order flow imbalance, order book depth change rate, and tick-by-tick directional ratio. Order flow imbalance can be quantified by calculating the ratio of the difference between active buy and sell volumes to the total trading volume within a sliding window; this indicator reflects the immediate comparison of buying and selling forces in the market. The order book depth change rate refers to the rate of change in the number of buy and sell orders at the top price level between the start and end of the window; it reflects the immediate supply and demand changes in market liquidity. The tick-by-tick directional ratio refers to the proportion of trades with prices higher than the previous trade within the window, used to measure the micro-directionality of price changes. These micro-structural features are calculated independently and in parallel within each sliding window, thereby generating a feature vector for each time granularity.

[0053] S403, Real-time monitoring of global market status parameters.

[0054] The global market state parameter is a macroeconomic indicator describing the current operating state of the entire financial market or a specific sector. It includes overall market volatility and liquidity indicators. Overall market volatility can be approximated by calculating the realized volatility of major market indices (such as the CSI 300 Index and the S&P 500 Index) over the most recent short period (e.g., the past 30 minutes). Liquidity indicators can be measured by calculating the median bid-ask spread, volume-weighted average bid-ask spread, or market depth index. These parameters are independent of individual financial instruments and reflect the overall "temperature" and "friction" of the market.

[0055] S404. Based on global market state parameters, dynamically calculate the current weight coefficient of each sliding window in the window pool using a preset adaptive weight function.

[0056] Specifically, the adaptive weighting function is a mathematical mapping that adjusts the importance of each window in real time based on global market state parameters. Its core logic is as follows: when overall market volatility increases, indicating a period of sharp market change, the instantaneous information captured by short-term windows (such as 1-minute or 5-minute windows) is more crucial for event identification; therefore, the adaptive weighting function increases the weight coefficients of these short-term windows. Conversely, when market volatility is stable, the weight of medium- to long-term windows is appropriately increased. When market liquidity indicators show decreased liquidity (such as widening bid-ask spreads), indicating increased market transaction costs and potentially amplified price shocks, the persistent trend information reflected by medium-term windows may be more reliable than instantaneous signals; therefore, the adaptive weighting function increases the weight coefficients of medium-term windows. Through this dynamic adjustment, the system can adaptively focus on the most relevant timescale information under different market conditions, avoiding information distortion or lag problems that may occur with fixed-weight schemes when market states change.

[0057] S405. Based on the current weight coefficients, the microstructural features of the market data output by each sliding window are weighted and fused to generate a multi-dimensional dynamic indicator vector representing the current market dynamics.

[0058] Specifically, weighted fusion refers to multiplying the feature vector of each window calculated in S402 by its current weight coefficient obtained in step S404, and then concatenating or summing all the weighted feature vectors to generate a comprehensive, multi-dimensional market dynamic indicator vector. For example, if the window pool contains 5 windows, and each window calculates 3 microstructure features, then the final generated market dynamic indicator vector may be a 15-dimensional vector. This vector not only contains market microstructure information at different time scales, but its weights have also been dynamically adjusted according to the current macroeconomic state of the market. Therefore, this vector can more comprehensively and accurately represent the dynamic characteristics of the target financial product in the current market environment, providing high-quality, high-information-density input for subsequent adaptive event recognition models.

[0059] Based on the above technical solution, this embodiment establishes a multi-scale, adaptive dynamic market indicator generation mechanism. This mechanism captures information across multiple time scales through a heterogeneous window pool, delves into the market's internal structure through parallel computation of microstructural features, perceives the macro environment by monitoring the overall market state, and achieves dynamic focusing through an adaptive weighting function. This design enables the generated dynamic market indicator vector to better adapt to rapidly changing market conditions, improving the accuracy and robustness of subsequent event identification and laying a solid data foundation for generating timely and accurate investment information.

[0060] Example 4, based on Example 1, such as Figure 3 As shown in the figure, this embodiment elaborates in detail on the specific construction and training method of the pre-trained adaptive event recognition model in S203. This model aims to establish an event recognition mechanism that can adapt to different market conditions and accurately identify specific structural behaviors or abnormal states in the financial market, including the following steps: S501. Organize the fused market dynamic indicator vectors generated by continuous time steps into a market dynamic indicator sequence in chronological order.

[0061] Specifically, in Example 3, a multi-dimensional market dynamic indicator vector is generated for each time step through heterogeneous sliding time window pooling and adaptive weight fusion. This step arranges these temporally continuous vectors according to their generation timestamps to form a time series. For example, if a fused market dynamic indicator vector is generated every minute, then a sequence containing 60 vectors will be formed over the past hour. This sequence fully records the dynamic evolution of the target financial product's microstructural characteristics over a period of time as the macroeconomic state of the market changes, providing basic data for the model to capture time-series dependencies.

[0062] S502. Utilize multi-source heterogeneous external data to synthesize a multi-dimensional soft label for each time segment in the market dynamic indicator sequence.

[0063] The multidimensional soft labels include event type, event intensity confidence level, and event duration estimate.

[0064] Specifically, to train the model to recognize events, labels need to be provided for the training data. Traditional hard labels (such as "yes / no" events) are too simplistic and fail to reflect the intensity and duration of events. This step employs a soft labeling strategy, utilizing multi-source heterogeneous external data to synthesize richer supervisory signals. This external data can include: authoritative financial news and announcements (to confirm the event type and approximate timing), research reports from professional analysts (to help determine the event intensity), historical backtesting data (to estimate the typical duration of events), and market sentiment indices (to quantify the market reaction triggered by the event). For example, for a price breakout event, the soft label might include the event type "price breakout," an event intensity confidence level of 0.85 (based on news mention frequency and analyst opinion consistency), and an estimated event duration of 2 hours (based on the average duration of similar historical events). This multi-dimensional soft labeling provides the model with more nuanced learning objectives that better reflect financial realities.

[0065] S503. Construct a dual-channel neural network model.

[0066] The dual channels include the main channel and the market status perception channel.

[0067] Specifically, the backbone channel processes dynamic market indicator sequences to extract time-series features. It can employ architectures adept at handling sequential data, such as Recurrent Neural Networks (RNNs), Long Short-Term Memory Networks (LSTMs), or Transformers. Its task is to extract deep time-series features from continuous dynamic market indicators and identify potential market patterns. The market state awareness channel receives global market state parameters and generates dynamic gating signals. This channel can be a relatively simple fully connected network or convolutional network. Its task is to analyze the current macro-market environment and generate a dynamic gating signal. This dual-channel design decouples micro-level target market pattern recognition from macro-level market environment awareness, allowing the model to focus on information at different levels. The dynamic gating signal is applied to the internal feature representation of the backbone channel for adaptive focusing on market patterns under different market conditions.

[0068] S504. Based on multidimensional soft labels and composite loss functions, the dual-channel neural network model is trained end-to-end to obtain a pre-trained adaptive event recognition model.

[0069] Specifically, the model is trained end-to-end, meaning it takes the original market dynamic indicator sequence and global market state parameters as input and directly outputs the event identification result. The training goal is to make the model's output as close as possible to the multidimensional soft labels synthesized in step S502. To this end, a composite loss function needs to be designed, which typically consists of three weighted parts: event type classification loss (e.g., cross-entropy loss, used to optimize the accuracy of event type prediction), event intensity regression loss (e.g., mean squared error loss, used to optimize the prediction accuracy of event intensity confidence), and event duration prediction loss (e.g., another mean squared error loss, used to optimize the estimation of event duration). By minimizing this composite loss function, the model can simultaneously learn to accurately determine the event type, assess the event intensity, and estimate the event duration.

[0070] Based on the above technical solution, this embodiment constructs an advanced and robust adaptive event recognition model. This model separates the processing of micro-modes from the processing of the macro-environment through a dual-channel architecture, achieves adaptive focusing on changes in market state through a dynamic gating mechanism, and undergoes refined training using multi-dimensional soft labels and a composite loss function. This design enables the model not only to identify "what" the event is, but also to assess "how strong" and "how long it might last," significantly improving the accuracy and information content of event recognition, laying a solid technical foundation for subsequently generating professional and timely investment consulting content.

[0071] Example 5, based on Example 1, elaborates on the specific implementation of step S204, namely, generating quantitative signal data based on the financial target corresponding to the market event data, combined with historical market data, a quantitative factor library, and strategy rules. This step aims to transform the identified market events into structured, interpretable, and executable quantitative investment analysis signals, including the following steps: S601. Based on the financial target, retrieve historical market data within a preset time window from the historical market database, and retrieve a set of quantitative factors that match the event type from the quantitative factor library.

[0072] Specifically, when step S203 identifies a market event (e.g., a stock experiencing a "price breakout"), the system locks the associated financial instrument (e.g., stock code 600XXX). Subsequently, the system retrieves historical market data for this financial instrument within a preset time window (e.g., the past 20 trading days) from the historical market database. This data may include daily opening price, closing price, highest price, lowest price, trading volume, and turnover. Simultaneously, the system retrieves a set of matching quantitative factors from the quantitative factor library based on the event type. The quantitative factor library is a pre-built set of factors containing multiple dimensions of financial analysis.

[0073] S602, based on a pre-configured strategy rule engine, performs multi-dimensional fusion analysis of historical market data, quantitative factors, and market event data.

[0074] The strategy rule engine includes condition-action mapping rules, which are used to determine the trading logic based on event type and factor combination.

[0075] Specifically, the strategy rule engine is the core processing unit of this step, containing a series of pre-configured "condition-action" mapping rules. These rules define what analytical actions should be performed under what market conditions or event contexts. For example, a rule might be: "If the event type is 'price breakout,' and the breakout amplitude exceeds 1.5 times the average true range of the past 20 days, while the trading volume amplification exceeds 2 times, then trigger the 'trend enhancement' analysis action." The specific logic of multi-dimensional fusion analysis can include: calculating the factor values ​​at the current moment based on historical market data and retrieved quantitative factors, such as the current RSI value and the current trading volume relative to the historical average. These factor values, along with market event data (such as event strength and event direction), are input into the strategy rule engine. The engine performs matching and inference according to the preset rules. The inference method can be rule-based logical judgment or prediction based on machine learning models (such as gradient boosting trees and neural networks).

[0076] S603. Based on the results of multidimensional fusion analysis, generate structured quantized signal data.

[0077] Specifically, quantitative signal data is a structured object whose content can include: signal direction, representing a long or short operation suggestion for a financial target, such as "bullish," "bearish," or "neutral"; signal strength, calculated based on a weighted average of factor exposure and event significance; confidence level, determined by a weighted average of event identification confidence and factor stability, indicating the credibility of the signal; and the basis for signal generation, recording the triggered strategy rule number, the name of the key quantitative factor involved in the calculation, and the type of market event, providing a traceable and interpretable basis for the signal.

[0078] In a preferred embodiment, when multidimensional fusion analysis generates multiple candidate quantization signals—for example, when different strategy rules give different directional judgments for the same event—the system performs further conflict handling. Specifically, it acquires the historical win rate, maximum drawdown, signal decay period, and current factor effectiveness score of the strategy corresponding to each candidate quantization signal under the same market conditions, and calculates the signal priority based on these indicators. Candidate quantization signals with a priority lower than a preset threshold are used as auxiliary reference signals and not as the primary quantization signals. Only when multiple candidate quantization signals have consistent directions and their confidence levels meet preset requirements does the system generate the primary quantization signal data. This conflict detection and sorting mechanism ensures that the final output quantization signal has higher reliability and consistency.

[0079] Based on the above technical solution, this embodiment establishes a multi-dimensional fusion quantitative signal generation mechanism based on a strategy rule engine. This mechanism extracts relevant information from historical data and a quantitative factor library, performs in-depth analysis using pre-configured rules or models, and ultimately generates structured, interpretable signal data. This design not only provides professional quantitative analysis basis for investment consulting but also enhances the transparency and credibility of the entire system through traceable signal generation evidence.

[0080] Example 6, based on Example 1, elaborates in detail on the specific implementation of step S205, namely, evaluating potential causal effects and generating consulting task data based on user investment context vectors, market event data, and quantitative signal data using a causal inference model, including the following steps: S701. Construct a causal inference model. This model uses the user's investment context vector as a condition variable and models market event data and quantitative signal data as intervention variables respectively.

[0081] Specifically, the causal inference model is the core analytical tool in this step, and its goal is to assess the causal impact of an "intervention" (i.e., the occurrence of a market event or quantitative signal) on an "outcome" (i.e., the return distribution of a user's portfolio). In this embodiment, the model can be constructed based on a potential outcome framework or a structural causal model.

[0082] S702 uses counterfactual reasoning to calculate the difference in the distribution of user portfolio returns with and without the intervention, and assesses the potential causal effect.

[0083] Specifically, counterfactual reasoning is the core logic of causal inference, attempting to answer questions like, "What would have happened if this event hadn't occurred?" In practical calculations, since only one scenario—either the intervention occurred or it didn't—can be observed, statistical or machine learning methods are needed to estimate the other, unobserved counterfactual outcome. For example, a feasible approach is to use a machine learning-based counterfactual prediction model. This model is first trained on historical data, learning the complex relationship between portfolio return distribution and market events and quantitative signals under given user context and various market conditions. When assessing the impact of a new event on the current user, the model makes two predictions: one inputting the current user context and the actual event / signal (intervention scenario), predicting the portfolio's return distribution; and another inputting the same user context but setting the event / signal variable to "not occurred" or "neutral" (counterfactual scenario), predicting a different return distribution. By comparing the difference between these two predicted distributions, the potential causal effect of the event or signal on the user's portfolio can be quantified.

[0084] S703. Based on the assessment results of potential causal effects and combined with preset consultation trigger rules, automatically generate structured consultation task data.

[0085] Specifically, the assessment result of potential causal effects is a vector or report containing multiple dimensions of impact measurement. The system needs to determine, based on these quantitative results, whether and how to generate consultation for the user. This relies on pre-defined consultation trigger rules. Consultation trigger rules can be a series of logical conditions. Based on the above technical solution, this embodiment establishes a consulting task generation mechanism based on causal inference. This mechanism, by constructing causal models, conducting counterfactual reasoning, and evaluating multi-dimensional effects, can more accurately quantify the actual impact of market events and signals on a specific user's investment portfolio, and generate consulting tasks supported by causal logic. Compared to traditional methods based solely on correlation or simple rule matching, this causal analysis-based approach can generate more targeted, interpretable, and scientific investment advice, effectively improving the quality and credibility of personalized consulting content.

[0086] Example 7, based on Example 1, elaborates on the specific implementation of step S207, namely, verifying the investment consulting text, including the following steps: Step S801: Perform multi-dimensional automatic verification on the investment consultation text.

[0087] Specifically, multi-dimensional automatic verification refers to the automated inspection of text content from multiple independent and complementary dimensions. These dimensions include compliance verification, consistency verification, data timeliness verification, and risk adaptation verification.

[0088] Step S8011: Perform compliance verification.

[0089] Specifically, compliance verification is based on a pre-built knowledge base of financial regulatory rules to detect whether investment advisory documents contain illegal language, profit promises, or misleading statements. The knowledge base of financial regulatory rules is a structured database that stores prohibited clauses and compliance requirements from laws and regulations such as the Securities Law, the Fund Law, and the Measures for the Administration of Investment Advisors, as well as industry self-regulatory rules.

[0090] Step S8012: Perform a consistency check.

[0091] Specifically, consistency verification compares the recommendations in the investment advisory text with the causal effect summaries and recommended actions in the original advisory task data to verify their logical consistency. This ensures that the final generated natural language text accurately and unambiguously reflects these structured analytical conclusions and recommendations. The verification engine extracts key conclusions and recommendations from the text and performs semantic comparison with corresponding fields in the advisory task data.

[0092] Step S8013: Perform data timeliness verification.

[0093] Specifically, data timeliness verification checks whether the interval between the timestamps of market events and the generation time of quantitative signals cited in investment advisory texts and the current system time is within a preset valid window. Financial markets are constantly changing, and advice based on outdated data may have lost its value or even be misleading. The preset valid window is a time threshold set according to data type and market volatility.

[0094] Step S8014: Perform risk adaptation verification.

[0095] Specifically, risk fit verification determines whether the risk level suggested in the investment advice text exceeds the user's acceptable range, based on the user's risk preference level in the investment context vector. The user's investment context vector contains the user's risk tolerance level, which defines the maximum level of risk the user can accept in investments. The actions suggested in the investment advice text also implicitly contain a corresponding risk level. The logic of risk fit verification is to match the risk level of the suggested actions in the text with the user's risk tolerance level.

[0096] Step S802: When all multi-dimensional automatic verifications pass, the investment consultation text is encrypted and pushed to the user terminal.

[0097] Step S803: If any verification item fails, the investment consultation text is corrected according to the failed verification item, and the corrected investment consultation text is re-verified.

[0098] Based on the above technical solutions, this embodiment establishes a rigorous and automated investment advisory text verification mechanism. This mechanism comprehensively covers key quality dimensions such as compliance, logic, timeliness, and risk through multi-dimensional parallel verification; ensures transmission security through encrypted push notifications; and improves the system's robustness and output success rate through a closed-loop correction mechanism. This design significantly reduces the probability of factual errors, logical contradictions, outdated data, or risk mismatches in intelligently generated content, thereby improving the security and reliability of the entire internet-based quantitative investment advisory generation system and enhancing users' trust in the platform's output content.

Claims

1. A method for generating quantitative investment information based on real-time market data, characterized in that: include: Collect real-time market data from multiple sources and build a real-time market data pool; Acquire basic and behavioral data of target users to generate user investment context vectors; the basic data includes user attribute information, including invested financial products, holding data, self-selected targets data, and risk tolerance level; the behavioral data includes user operation records on the platform. Based on the real-time market data pool, dynamic market indicator vectors at different time granularities are calculated and input into a pre-trained adaptive event recognition model to identify and generate market event data. Based on the financial instruments corresponding to the market event data, quantitative signal data is generated by combining historical market data, quantitative factor library and strategy rules; Based on the user investment context vector, the market event data, and the quantitative signal data, the potential causal effects of the user's investment portfolio are assessed, and consulting task data is generated. Based on the aforementioned consulting task data, a consulting generation template and a language generation model are invoked to generate investment consulting text; The investment advice text is verified, and the verified investment advice text is output to the user terminal.

2. The method for generating internet-based quantitative investment information based on real-time market data linkage according to claim 1, characterized in that, The process of acquiring basic and behavioral data of the target user and generating a user investment context vector includes: Based on the behavioral data, a dynamic intent map representing users' short-term investment intentions and their correlations is constructed through natural language processing and graph neural networks; Based on the attribute information and real-time market data, a real-time risk status indicator representing the user's current actual risk exposure level is calculated. The basic data, the dynamic intent graph, and the real-time risk status indicators are used as multimodal inputs and processed through a cross-modal attention fusion network to generate a user investment context vector.

3. The method for generating internet-based quantitative investment information based on real-time market data linkage according to claim 2, characterized in that, The method for obtaining the risk status indicators includes: Using the tradable financial products in the attribute information as nodes, and calculating the dynamic correlation between any two nodes based on real-time market data as edge weights, a holding-watching coupled risk network is constructed. For each node in the coupled risk network, an implicit volatility surface is fitted to its corresponding real-time market data, and a multi-dimensional risk feature vector is extracted from it. Based on the topology of the coupled risk network, the multidimensional risk feature vectors of each node are weighted and aggregated to obtain the overall risk features of the user combination. The overall risk characteristics are compared with the preset risk preference vector of the risk tolerance level to generate a real-time risk status index that includes absolute risk level and relative risk fit.

4. The method for generating internet-based quantitative investment information based on real-time market data linkage according to claim 1, characterized in that, The analysis method for the aforementioned market dynamic indicator vector includes: Construct a window pool containing multiple heterogeneous sliding windows with different time granularities; wherein the time length of each window is not uniformly distributed and contains at least one asymmetric window; For the real-time market data stream of the target financial product, a set of market microstructure features are calculated in parallel within each sliding window. The market microstructure features include at least two of the following: order flow imbalance, bid-ask depth change rate, and tick-by-tick directional ratio. Real-time monitoring of global market status parameters, including overall market volatility and liquidity indicators; Based on the global market state parameters, the current weight coefficient of each sliding window in the window pool is dynamically calculated through a preset adaptive weight function; Based on the current weighting coefficients, the microstructural features of the market data output from each sliding window are weighted and fused to generate a multidimensional dynamic indicator vector representing the current market dynamics.

5. The method for generating internet-based quantitative investment information based on real-time market data linkage according to claim 1, characterized in that, The pre-trained adaptive event recognition model includes: The fused market dynamic indicator vectors generated by continuous time steps are organized into a market dynamic indicator sequence in chronological order. Using multi-source heterogeneous external data, a multi-dimensional soft label is synthesized for each time segment in the market dynamic indicator sequence. The multi-dimensional soft label includes event type, event intensity confidence, and event duration estimation. A dual-channel neural network model is constructed; wherein the dual channels include a backbone channel and a market state perception channel; the backbone channel is used to process the dynamic market indicator sequence to extract time-series features, and the market state perception channel is used to receive global market state parameters and generate dynamic gating signals; Based on the multidimensional soft labels and the composite loss function, the dual-channel neural network model is trained end-to-end to obtain a pre-trained adaptive event recognition model; the composite loss function includes event type classification loss, event intensity regression loss, and event duration prediction loss.

6. The method for generating internet-based quantitative investment information based on real-time market data linkage according to claim 1, characterized in that, The method for generating the quantized signal data includes: Based on the financial target, retrieve historical market data within a preset time window from the historical market database, and retrieve a set of quantitative factors that match the event type from the quantitative factor library; Based on a pre-configured strategy rule engine, the historical market data, the quantitative factors, and the market event data are fused and analyzed in multiple dimensions; wherein, the strategy rule engine includes condition-action mapping rules, which are used to determine the trading logic based on the event type and factor combination; Based on the results of the multidimensional fusion analysis, structured quantized signal data is generated.

7. The method for generating internet-based quantitative investment information based on real-time market data linkage according to claim 1, characterized in that, The method for obtaining the consultation task data includes: A causal inference model is constructed, which uses the user's investment context vector as a condition variable and models the market event data and quantitative signal data as intervention variables respectively. The difference in the distribution of user portfolio returns with or without the intervention is calculated by counterfactual reasoning. The potential causal effect of the market events and quantitative signals on the user portfolio is evaluated. The potential causal effect includes the impact on portfolio volatility, maximum drawdown, expected return and position correlation. Based on the assessment results of the potential causal effects and combined with the preset consultation triggering rules, structured consultation task data is automatically generated.

8. The method for generating internet-based quantitative investment information based on real-time market data linkage according to claim 1, characterized in that, The verification of the investment advisory text includes: Perform multi-dimensional automatic verification on the investment consultation text, the multi-dimensional automatic verification including: Compliance verification: Based on a pre-built knowledge base of financial regulatory rules, detect whether there are any illegal terms, profit promises or misleading statements in the text; Consistency verification: Compare the suggested content in the investment advice text with the causal effect summary and recommended actions in the original advice task data to verify logical consistency; Data timeliness verification: Check whether the interval between the market event timestamps and quantitative signal generation times cited in the investment consultation text and the current system time is within a preset valid window; Risk Adaptability Verification: Based on the risk preference level in the user's investment context vector, determine whether the operational risk level suggested in the investment advice text exceeds the user's acceptable range; When all multi-dimensional automatic verifications pass, the investment consultation text will be encrypted and pushed to the user's terminal. If any verification item fails, the investment advisory text shall be corrected according to the failed verification item, and the corrected investment advisory text shall be re-verified.

9. The method for generating internet-based quantitative investment information based on real-time market data linkage according to claim 1, characterized in that, The construction of the real-time market data pool includes: Subscribe to or poll to collect raw real-time market data from multiple market data sources for at least one financial instrument; The raw real-time market data is mapped and aligned with timestamps to form preliminary standardized data; Outlier filtering is performed on the preliminary standardized data, and cross-validation is performed on the preliminary standardized data from different market data sources that are for the same financial instrument and the same market data time to identify conflicting data. Based on the historical performance data of each market data source, a dynamic evaluation model for market data source quality is constructed. The historical performance data includes the historical delay distribution and historical price deviation distribution of each market data source under different preset market conditions. The dynamic evaluation model for market data source quality is configured to output the dynamic credibility weight of each market data source in real time according to the current market condition. Using the dynamic credibility weight, the conflicting data is adaptively weighted and fused to generate fused market data; The fused market data, the preliminary standardized data that has passed cross-validation and is conflict-free, and the data marked as abnormal or pending review are all stored in the real-time market data pool, and each piece of data in the pool is associated with its source information and fusion identifier.

10. An internet-based quantitative investment advisory generation system based on real-time market data linkage, operating based on the internet-based quantitative investment advisory generation method based on real-time market data linkage as described in any one of claims 1-9, characterized in that... It includes an acquisition module, an identification module, a quantification module, a consultation module, and a verification module; The acquisition module is used to collect real-time market data from multiple sources and construct a real-time market data pool. Acquire basic and behavioral data of target users to generate user investment context vectors; The identification module is used to calculate dynamic indicator vectors of market data at different time granularities based on the real-time market data pool, and input them into a pre-trained adaptive event recognition model to identify and generate market event data. The quantitative module is used to generate quantitative signal data based on the financial target corresponding to the market event data, combined with historical market data, a quantitative factor library, and strategy rules. The consultation module is used to assess the potential causal effects of a user's investment portfolio based on the user's investment context vector, the market event data, and the quantitative signal data, and to generate consultation task data. Based on the aforementioned consulting task data, a consulting generation template and a language generation model are invoked to generate investment consulting text; The verification module is used to verify the investment consultation text and output the verified investment consultation text to the user terminal.

Images