A large model-based personalized financial service recommendation method and system

Abstract

The application relates to the technical field of financial technology, and in particular to a personalized financial service recommendation method and system based on a large model. The method comprises the following steps: obtaining structured financial behavior data and unstructured financial expression data of a target customer, inputting a large model recommendation model to generate a customer portrait feature vector and an initial recommendation set; obtaining and analyzing external market information to generate market event records, update the risk degree and attractiveness of financial products; calculating a calibrated recommendation score according to the original recommendation score, the risk degree and the attractiveness, filtering and reordering the initial recommendation set to obtain a target recommendation set. The scheme can alleviate the problems of customer portrait distortion and disconnection between the recommendation result and the market reality caused by improper processing of unstructured financial expression data.

Description

Technical Field

[0001] This application relates to the field of financial technology, and in particular to a method and system for recommending personalized financial services based on a large model. Background Technology

[0002] Existing financial service recommendation methods based on large-scale intelligent systems perform well when processing traditional structured customer transaction information, but they have significant limitations when facing younger customer groups. These customers exhibit highly non-linear and fragmented financial behaviors, making their investment intentions difficult to accurately capture using traditional structured data. More importantly, these customers frequently exchange financial opinions using unstructured data containing internet slang, emoticons, and multilingual expressions on social media and technology forums. Existing systems have hidden flaws in their data preprocessing stages when handling such complex information. When processing text containing specific combinations of emoticons or rare multilingual characters, data alignment errors or partial loss of feature vector data can occur. This intermittent data contamination directly undermines the accuracy of customer profile generation and update logic, leading to lags and biases in customer profiles. As a result, recommendation systems frequently push inappropriate financial products to customers, severely reducing the surprise and relevance of recommendations. Faced with these low-quality recommendations, customers often choose silence rather than proactive feedback, forming a silent negative feedback loop that prevents the system from effectively learning and optimizing from actual interactions. Meanwhile, existing systems lack the ability to perceive and dynamically adjust external market information in real time, making it difficult to adaptively adjust recommendation strategies based on real-time market dynamics such as breaking news and policy changes. This results in a serious disconnect between recommendation results and actual market needs, ultimately affecting the service quality and market competitiveness of financial institutions. Therefore, existing technologies urgently need improvement to address these issues. Summary of the Invention

[0003] This application discloses a personalized financial service recommendation method and system based on a large model, aiming to solve the technical problems in the prior art caused by improper processing of unstructured financial data, resulting in distorted customer profiles and a disconnect between recommendation results and market reality.

[0004] The technical solution of this application is as follows: Firstly, this application discloses a personalized financial service recommendation method based on a large model, including: Acquire customer characteristic data of target customers, including structured financial behavior data and unstructured financial expression data of target customers; Customer feature data is input into the large-scale recommendation model to generate a customer profile feature vector for the target customer. Based on the matching relationship between the customer profile feature vector and the product feature vectors of each financial product in the financial product library, an initial recommendation set is generated. The large-scale recommendation model is a model used for semantic understanding, behavioral analysis and preference extraction of customer feature data. The initial recommendation set includes at least one financial product and the original recommendation score corresponding to the financial product. We obtain external market information through external data interfaces, including macroeconomic data, financial regulatory policy information, and breaking news events. Perform event analysis on external market information to generate at least one market event record. Each market event record includes fields for the affected object, the direction of risk impact, the direction of attractiveness impact, the intensity of risk impact, and the intensity of attractiveness impact. Based on the fields of affected object, risk impact direction, attractiveness impact direction, risk impact intensity, and attractiveness impact intensity in the market event record, determine the risk level adjustment and attractiveness adjustment for the affected object. Then, based on the risk level adjustment and attractiveness adjustment, update the market status parameters corresponding to the affected object in the real-time market status parameter table. The real-time market status parameter table is used to store the market status parameters corresponding to financial products or financial product categories. The market status parameters include risk level and attractiveness. For the financial products in the initial recommendation set, the corresponding risk level and attractiveness are read from the real-time market status parameter table, and the calibration recommendation score is calculated based on the original recommendation score, risk level and attractiveness. Based on risk level, attractiveness, and calibration recommendation score, the initial recommendation set is modified to obtain the target recommendation set. The modification includes filtering financial products with risk levels higher than a preset risk threshold and / or filtering financial products with attractiveness lower than a preset attractiveness threshold, and re-sorting the unfiltered financial products according to the calibration recommendation score.

[0005] Secondly, this application also discloses a personalized financial service recommendation system based on a large model, including: The customer characteristic data acquisition module is used to acquire customer characteristic data of target customers, including structured financial behavior data and unstructured financial expression data of target customers. The initial recommendation set generation module is used to input customer feature data into the large model recommendation model, generate customer profile feature vectors of target customers, and generate an initial recommendation set based on the matching relationship between the customer profile feature vectors and the product feature vectors of each financial product in the financial product library. The large model recommendation model is a model used for semantic understanding, behavioral analysis and preference extraction of customer feature data. The initial recommendation set includes at least one financial product and the original recommendation score corresponding to the financial product. The external market information acquisition module is used to acquire external market information through external data interfaces. External market information includes macroeconomic data, financial regulatory policy information, and breaking news events. The market event analysis module is used to analyze external market information and generate at least one market event record. Each market event record includes fields for the affected object, the direction of risk impact, the direction of attractiveness impact, the intensity of risk impact, and the intensity of attractiveness impact. The market status parameter adjustment module is used to determine the risk level adjustment amount and attractiveness adjustment amount corresponding to the affected object based on the affected object field, risk impact direction field, attractiveness impact direction field, risk impact intensity field, and attractiveness impact intensity field in the market event record. Based on the risk level adjustment amount and attractiveness adjustment amount, it updates the market status parameters corresponding to the affected object in the real-time market status parameter table. The real-time market status parameter table is used to store the market status parameters corresponding to financial products or financial product categories. The market status parameters include risk level and attractiveness. The calibration recommendation score calculation module is used to read the corresponding risk level and attractiveness from the real-time market status parameter table for financial products in the initial recommendation set, and calculate the calibration recommendation score based on the original recommendation score, risk level and attractiveness. The recommendation set correction module is used to correct the initial recommendation set based on the risk level, attractiveness, and calibration recommendation score to obtain the target recommendation set. The correction includes: filtering financial products with a risk level higher than a preset risk threshold, and / or filtering financial products with an attractiveness lower than a preset attractiveness threshold, and reordering the unfiltered financial products according to the calibration recommendation score.

[0006] Beneficial Effects: This application constructs a personalized financial service recommendation scheme based on a large-scale model. By utilizing this model for deep semantic understanding, behavioral analysis, and preference extraction of customer structured and unstructured feature data, it effectively overcomes the customer profile distortion problem caused by deficiencies in unstructured data processing in existing technologies. Simultaneously, by acquiring and analyzing external market information in real time, it dynamically adjusts the risk level and attractiveness parameters of financial products, achieving an adaptive response of the recommendation strategy to changes in the market environment. This method not only restores the effectiveness of the customer feedback loop, ensuring the recommendation system can continuously optimize based on accurate user intent, but also significantly improves the foresight and risk management capabilities of the recommendation results through real-time perception and quantitative analysis of market events, ultimately achieving a dual improvement in the quality of financial service recommendations and customer satisfaction. Attached Figure Description

[0007] Figure 1 This is a flowchart illustrating a personalized financial service recommendation method based on a large model, which is provided in this application.

[0008] Figure 2 A flowchart of a personalized financial service recommendation system based on a large model is provided for this application.

[0009] In the diagram: 1. Customer feature data acquisition module; 2. Initial recommendation set generation module; 3. External market information acquisition module; 4. Market event analysis module; 5. Market status parameter adjustment module; 6. Calibration recommendation score calculation module; 7. Recommendation set correction module. Detailed Implementation

[0010] The embodiments of this application are described below with reference to the accompanying drawings. The described embodiments are only some embodiments of this application and are not intended to limit the scope of protection of this application. In the description of this application, the terms "first," "second," etc., are used only for distinguishing descriptions and should not be construed as indicating or implying relative importance.

[0011] For ease of understanding, structured financial behavioral data refers to customer financial activity records with predefined fields; unstructured financial expression data refers to financial-related text containing natural language or symbolic expressions; large-scale recommendation models are used to perform semantic understanding, behavioral analysis, and preference extraction on customer feature data; market event records are used to describe the impact of external events on the risk level and attractiveness of financial products or financial product categories; and real-time market state parameter tables are used to dynamically store the risk level and attractiveness corresponding to financial products or financial product categories.

[0012] This application provides a personalized financial service recommendation method based on a large model, which includes the following steps: S1000: Obtain customer characteristic data of the target customer, which includes structured financial behavior data and unstructured financial expression data of the target customer.

[0013] In practice, structured financial behavior data can be obtained from the financial institution's internal data warehouse, including customer account balance change records, transaction frequency, historical holdings list, subscription and redemption operation logs, etc.

[0014] In practice, the system can extract transaction data from the core transaction system, and after deduplication, missing value processing, time format standardization and integrity verification, it forms customer behavior data containing customer ID, transaction timestamp, transaction type, transaction amount, product code and transaction channel.

[0015] Acquiring unstructured financial data involves interface integration with external social media platforms. In practice, a distributed web crawler system (based on the Scrapy or Apache Nutch framework) can be deployed to monitor specific financial keywords (such as specific token names "BTC" or "ETH", or decentralized finance protocol identifiers "Uniswap" or "Compound"). Provided that interface authorization is obtained and data usage rules are followed, text content containing these keywords can be captured in real time.

[0016] For younger investors, special attention should be paid to text snippets containing emojis (such as rocket emojis, moon emojis, diamond hand emojis, etc.), non-standard abbreviations (such as "FOMO" for fear of missing out and "HODL" for long-term holding), and multilingual expressions.

[0017] S2000: Input customer feature data into the large-scale recommendation model to generate customer profile feature vectors for target customers, and generate an initial recommendation set based on the matching relationship between the customer profile feature vectors and the product feature vectors of each financial product in the financial product library.

[0018] The large-scale recommendation model is used for semantic understanding, behavioral analysis, and preference extraction of customer feature data. The initial recommendation set includes at least one financial product and the original recommendation score corresponding to the financial product.

[0019] In this step, the large-scale recommendation model can be a pre-trained language model based on the Transformer architecture (such as BERT-large or FinBERT in the financial field), which is then fine-tuned with specific corpora in the financial field and deployed on the recommendation server.

[0020] Subsequently, for the recommendation task, a supervised training dataset was constructed: customer feature data (including structured financial behavior data and unstructured financial expression data) was selected from historical recommendation records as input, the products that the customer finally purchased were used as positive sample labels, and the products that the customer browsed but did not purchase were used as negative sample labels. The model parameters were fine-tuned using contrastive learning.

[0021] The model first encodes the input customer feature data: for structured financial behavior data, discrete categorical features (such as product type preference) are transformed into 128-dimensional dense vectors through the embedding layer, and continuous numerical features (such as the mean of transaction amount and the standard deviation of holding days) are concatenated after Min-Max normalization; for unstructured text data, the model's self-attention mechanism is used to capture long-distance semantic dependencies and identify sentiment polarity, investment theme tendencies, and risk preference cues in the text.

[0022] The model output layer fuses and compresses various features into a fixed-dimensional customer profile feature vector.

[0023] The financial product database pre-stores product feature vectors for each financial product. These product feature vectors can be generated by encoding information such as product descriptions, historical return characteristics, and risk rating texts, and a mapping relationship is established with the corresponding financial product identifier.

[0024] The recommendation server calculates the cosine similarity between the customer profile feature vector and each product feature vector using the following formula: Sim = (Vc·Vp) / (||Vc|| × ||Vp||), where Sim is the similarity between the customer profile feature vector and the product feature vector, Vc is the customer profile feature vector, Vp is the product feature vector, Vc·Vp is the dot product of Vc and Vp, ||Vc|| is the vector norm of Vc, and ||Vp|| is the vector norm of Vp. When ||Vc|| or ||Vp|| is 0, the corresponding similarity can be set to 0, or the corresponding vector can be subjected to preset smoothing processing.

[0025] Based on the similarity scores, the top N (e.g., top 20) financial products are selected to form the initial recommendation set, where N is the number of candidate products in the initial recommendation set. Each selected financial product is associated with an original recommendation score, which is the result of the similarity calculation above, mapped to a value between 0 and 1 using the sigmoid function, representing the degree of match between the customer and the product.

[0026] S3000: Obtains external market information through external data interfaces, including macroeconomic data, financial regulatory policy information, and breaking news events.

[0027] External market information can be obtained through macroeconomic data interfaces, regulatory policy information interfaces, and breaking news event interfaces. Different collection frequencies can be set according to the information type. Furthermore, external market information includes at least one of macroeconomic data, financial regulatory policy information, breaking news event information, and emerging financial information source data. Emerging financial information source data includes the opinions of community opinion leaders.

[0028] S4000: Performs event analysis on external market information and generates at least one market event record. Each market event record includes fields for the affected object, the direction of risk impact, the direction of attractiveness impact, the intensity of risk impact, and the intensity of attractiveness impact.

[0029] The event parsing process transforms macroeconomic data, financial regulatory policy information, and breaking news events into structured market event records that can be accessed by real-time market state parameter tables. For unstructured news texts and policy texts, the event parsing process involves the structured extraction of text content using natural language processing techniques.

[0030] In practical implementation, Named Entity Recognition (NER) models (such as those based on the BERT-BiLSTM-CRF architecture) can be used to locate the names or categories of financial products mentioned in the text (such as "Bitcoin futures," "green bond funds," and "CSI 300 ETF"), and these can be filled into the "Affected Object" field. Sentiment analysis models or rule engines can be used to determine the direction of the event's impact on risk levels and investment attractiveness.

[0031] The impact of an event on the product's risk level and investment attractiveness can be determined through a sentiment analysis model or a pre-set rule engine. The sentiment analysis model can use a FinBERT-based text classification model to divide text into three categories: "positive," "neutral," and "negative." At the same time, an independent intent recognition model is trained to determine whether the text involves risk warnings (such as "default," "running away," or "regulatory crackdown") or positive stimuli (such as "technological breakthrough," "cooperation agreement," or "expected performance increase").

[0032] If the news involves tightening regulations (such as "the central bank summoned a trading platform"), the risk impact direction field is marked as "increased" and the attractiveness impact direction field is marked as "weakened"; if it is a technological breakthrough announcement (such as "a public chain completed a major upgrade, increasing TPS to 100,000"), the risk impact direction field can be marked as "decreased" (due to improved technological maturity reducing operational risks) and the attractiveness impact direction field is marked as "enhanced".

[0033] The risk impact strength field and the attractiveness impact strength field can be generated through a trained regression model. The model output is an integer value between 0 and 10, with larger values ​​indicating a more severe impact. For example, news about a country completely banning cryptocurrency trading might be assigned a risk impact strength value of 9, while news about a fund's fee adjustment might only be assigned a strength value of 2.

[0034] S5000: Based on the fields of the affected object, the direction of risk impact, the direction of attractiveness impact, the intensity of risk impact, and the intensity of attractiveness impact in the market event record, determine the risk level adjustment and attractiveness adjustment for the affected object, and update the market state parameters corresponding to the affected object in the real-time market state parameter table according to the risk level adjustment and attractiveness adjustment. The real-time market state parameter table is used to store the market state parameters corresponding to financial products or financial product categories. The market state parameters include risk level and attractiveness.

[0035] The real-time market status parameter table stores the current market status parameters corresponding to financial products or financial product categories. Each record corresponds to a financial product identifier or financial product category identifier and is associated with a risk level score and an attractiveness score. The risk level score ranges from 0 to 100, with a higher value indicating higher risk. The attractiveness score ranges from 0 to 100, with a higher value indicating stronger attractiveness.

[0036] During system initialization, the initial value of the risk level score can be set according to the basic risk level of the product. For example, the initial risk level of an R1 product is 20 points, and the initial risk level of an R5 product is 80 points. The initial value of the attractiveness score can also be set according to the recent return of the product. For example, the initial attractiveness of a product with a positive return in the past month is 60 points, and the initial attractiveness of a product with a negative return in the past month is 40 points.

[0037] After a new market event record is generated, the system calculates the risk level adjustment based on the risk impact direction and risk impact intensity fields in the market event record, and calculates the attractiveness adjustment based on the attractiveness impact direction and attractiveness impact intensity fields. The specific calculations can use the following linear adjustment formulas: ΔRisk = kRisk × b × Ir; ΔAttr = kAttr × b × Ia, where ΔRisk is the risk level adjustment, ΔAttr is the attractiveness adjustment, kRisk is the risk adjustment direction coefficient, kAttr is the attractiveness adjustment direction coefficient, b is the basic adjustment step size, Ir is the risk impact intensity value, and Ia is the attractiveness impact intensity value, with Ir and Ia both taking integer values ​​between 0 and 10.

[0038] In the above formula, when the direction of risk impact is "increase", kRisk = +1; when the direction of risk impact is "decrease", kRisk = -1; when the direction of risk impact is "maintain", kRisk = 0; when the direction of attractiveness impact is "enhance", kAttr = +1; when the direction of attractiveness impact is "weaken", kAttr = -1; when the direction of attractiveness impact is "maintain", kAttr = 0; the basic adjustment step size b can be 5 points or preset according to business rules.

[0039] During the update, the system determines the financial product or financial product category to be updated in the real-time market status parameter table based on the affected object field. If the affected object field corresponds to a specific financial product identifier, the market status parameter corresponding to that financial product is updated directly; if the affected object field corresponds to a financial product category, the market status parameter corresponding to the financial product belonging to that category is updated. Subsequently, the calculated risk level adjustment and attractiveness adjustment are accumulated to the current value of the corresponding record in the real-time market status parameter table, and the risk level score and attractiveness score can be updated according to Risk_new = min(100, max(0, Risk_old + ΔRisk)) and Attr_new = min(100, max(0, Attr_old + ΔAttr)), where Risk_new is the updated risk level score, Risk_old is the original risk level score, Attr_new is the updated attractiveness score, and Attr_old is the original attractiveness score. The min function is used to limit the upper limit of the score, and the max function is used to limit the lower limit of the score.

[0040] For example, if a cryptocurrency's current risk level score (Risk_old) is 60, the risk impact intensity value (Ir) of a sudden security vulnerability event is 7, the base adjustment step size (b) is 5, and the risk impact direction is "increased", then kRisk = +1, ΔRisk = 1×5×7 = 35, and the updated risk level score (Risk_new) = min(100, max(0, 60+35)) = 95, which is close to the risk limit and can trigger a high-risk warning.

[0041] S6000: For financial products in the initial recommendation set, read the corresponding risk level and attractiveness from the real-time market state parameter table, and calculate and calibrate the recommendation score based on the original recommendation score, risk level, and attractiveness.

[0042] The calibration recommendation score is calculated to integrate customer fit with market conditions. In practice, a multiplicative calibration formula can be used: Cm = P × (Attr / 100) × (1 - Risk / 100), where Cm is the calibration recommendation score, P is the original recommendation score, Attr is the attractiveness, and Risk is the risk level; P ranges from 0 to 1, and Attr and Risk both range from 0 to 100.

[0043] This formula makes the calibration recommendation score influenced by the original recommendation score, attractiveness, and risk level simultaneously: when the original recommendation score is high but the risk level is high or the attractiveness is low, the calibration recommendation score will decrease accordingly. For example, when the original recommendation score P is 0.9, the attractiveness Attr is 30, and the risk level Risk is 90, Cm = 0.9 × (30 / 100) × (1 - 90 / 100) = 0.027.

[0044] The system performs the above calculations for each financial product in the initial recommendation set to obtain a calibrated recommendation score that reflects the current market conditions. This calibrated recommendation score preserves the matching relationship between the customer profile feature vector and the product feature vector, while also incorporating risk level and attractiveness to calibrate the recommendation results according to market conditions.

[0045] S7000: Based on the risk level, attractiveness, and calibration recommendation score, the initial recommendation set is modified to obtain the target recommendation set. The modification includes: filtering financial products with a risk level higher than a preset risk threshold, and / or filtering financial products with an attractiveness lower than a preset attractiveness threshold, and re-sorting the unfiltered financial products according to the calibration recommendation score.

[0046] The preset risk and attraction thresholds can be dynamically set according to the customer's risk preference level. The system can classify customers' risk preferences into conservative, moderate, balanced, growth-oriented, and aggressive levels, and configure corresponding risk and attraction thresholds for different customer risk preference levels. For example, for conservative customers, the risk threshold can be set to 40 points and the attraction threshold can be set to 30 points; for aggressive customers, the risk threshold can be set to 80 points and the attraction threshold can be set to 20 points.

[0047] The system iterates through the initial recommendation set, removing financial products whose risk level is higher than the preset risk threshold or whose attractiveness is lower than the preset attractiveness threshold; the remaining financial products are sorted in descending order according to their calibration recommendation scores, and the top M financial products are selected to form the target recommendation set, where M is the number of financial products in the target recommendation set, and M can be preset according to the number of recommendations displayed or business rules.

[0048] Furthermore, external market information is analyzed to generate at least one market event record, including: S4100: Obtain the opinions of community opinion leaders from pre-defined emerging financial information sources, including at least one of the following: technology forums, developer communities, crypto asset-related mailing lists, and industry analyst information sources; S4200: Performs domain semantic analysis on the statements of community opinion leaders to identify technical terms, informal expressions and industry jargon involved in the statements, and determines the core viewpoints, sentiment tags, potential market impact tags and related financial products or related product categories corresponding to the statements based on the identification results. S4300: Obtain the influence score corresponding to the community opinion leader. The influence score is determined based on at least one of the following: the frequency of citation of the community opinion leader's remarks, the amount of content dissemination, the historical prediction accuracy, the amount of community interaction, and the activity level. S4400: Generate a community viewpoint market event record based on core viewpoints, sentiment tags, potential market impact tags, related financial products or related product categories, and influence scores to form one of the market event records. The impact object field of the community viewpoint market event record is determined based on the related financial products or related product categories. The risk impact direction field and attractiveness impact direction field of the community viewpoint market event record are determined based on sentiment tags and potential market impact tags. The risk impact strength field and attractiveness impact strength field of the community viewpoint market event record are determined based on the influence score.

[0049] Relying solely on traditional financial media to obtain external market information makes it difficult to capture the sentiment trends of emerging financial information sources such as technology forums and developer communities. Therefore, an event analysis mechanism can be set up for the remarks of community opinion leaders.

[0050] In practical implementation, pre-defined emerging financial information sources may include blockchain technology forums, developer communities, crypto asset-related mailing lists, industry analyst information sources, or independent analyst publishing channels. The system can obtain text content from the above information sources through data interfaces or web crawling technology, and then perform subsequent domain semantic analysis on the obtained text content.

[0051] Community opinion leaders can be determined based on a preset list or dissemination influence indicators. These dissemination influence indicators may include at least one of the following: number of followers, post interaction rate, number of citations, PageRank value, or HITS algorithm authority score. The PageRank value or HITS algorithm authority score is used to characterize a user's information dissemination influence and the degree to which their opinions are cited within emerging financial information sources.

[0052] After acquiring the opinions of community opinion leaders, domain semantic parsing needs to process non-standardized text. For example, for text containing "The code audit of this project has just been completed, there are no critical issues, so we can confidently go all in with the rocket emoji," the parsing system can identify "all in" as industry slang, "rocket emoji" as a bullish sentiment symbol, and "code audit" and "critical issues" as technical terms, and map them to standardized semantic tags. For example, the core viewpoint is "The project's security has been verified, and a buy recommendation is given," the sentiment tag is "positively bullish," the potential market impact tag is "may trigger increased buying," and the associated financial product is the specific financial product or category of financial products discussed in the text.

[0053] Influence score is calculated by comprehensively considering the frequency of citations, content dissemination, historical prediction accuracy, community interaction volume, and activity level. Historical prediction accuracy is determined by comparing past opinions of community opinion leaders with subsequent actual market trends. Specifically, it involves extracting statements containing price predictions or trend judgments within the statistical period, recording the statement's publication time (Tpred) and prediction direction, and obtaining market price data for a preset time period after Tpred. A prediction is considered accurate when the predicted direction aligns with the actual price change. Historical prediction accuracy PredAcc = AccurateNum / PredictNum × 100%, where PredAcc is the historical prediction accuracy, AccurateNum is the number of accurate predictions, PredictNum is the total number of valid predictions within the statistical period, and Tpred is the time when the community opinion leader published a statement containing price predictions or trend judgments.

[0054] These metrics are weighted and summed to obtain an influence score from 0 to 100. The calculation formula is: Inf = 0.30×QuoteFreq + 0.20×Spread + 0.40×PredAcc + 0.10×EngageAct, where Inf is the influence score, QuoteFreq is the frequency of citations normalized to the 0-100 range, Spread is the content dissemination volume normalized to the 0-100 range, PredAcc is the historical prediction accuracy, and EngageAct is the combined value of community interaction and activity normalized to the 0-100 range.

[0055] When generating community opinion market event records, the "Affected Object" field is determined based on the parsed related financial products or related product categories. The "Risk Impact Direction" and "Attractiveness Impact Direction" fields can be determined based on sentiment tags and potential market impact tags, and validated in conjunction with the core viewpoint: if the sentiment tag is "negative and bearish" and the potential market impact tag is "may trigger sell-offs," and the core viewpoint indicates a risk warning or negative evaluation, then the risk impact direction is set to "increased," and the attractiveness impact direction is set to "weakened." If the sentiment tag is "positive and bullish" and the potential market impact tag is "may trigger increased buying," and the core viewpoint indicates improved project security, technological advancements, or a buy recommendation, then the risk impact direction is set to "decreased," and the attractiveness impact direction is set to "enhanced."

[0056] The Risk Impact Strength and Attractiveness Impact Strength fields are non-negative numerical fields, determined based on the influence score. When these fields need to be consistent with the calculation formulas for subsequent risk level adjustments and attractiveness adjustments, the influence score can be normalized to the range of 0 to 10, i.e., Ir = Inf / 10, Ia = Inf / 10, where Ir is the risk impact strength value, Ia is the attractiveness impact strength value, and Inf is the influence score. For example, when a community opinion leader with an influence score of Inf of 80 publishes a strongly bearish opinion, a community opinion market event record can be generated with a risk impact direction of "increased," a risk impact strength value of Ir of 8, an attractiveness impact direction of "decreased," and an attractiveness impact strength value of Ia of 8.

[0057] Preferably, S4200 includes: S4210: Extract at least two types of semantic signals from the statements of community opinion leaders. Semantic signals include technical terms, text fragments, emojis, code snippets, non-standard abbreviations, or industry jargon. S4220: Identify information about financial entities, opinions, sentiments, and market impact related to financial products or categories of financial products in various semantic signals; S4230: Match financial entity information with financial product identifiers and / or financial product category identifiers in a financial product database to determine the associated financial product or associated product category corresponding to the statement; S4240: Based on the associated financial products or product categories, opinion information, sentiment information and market impact information corresponding to different semantic signals, calculate the correlation strength and consistency between different semantic signals to obtain semantic synergy results or semantic conflict results. S4250: When a semantic conflict result is obtained, the conflicting semantic signals are evaluated by weight according to the preset context knowledge base to obtain the weight value corresponding to the conflicting semantic signals. The preset context knowledge base includes at least one of financial community expression patterns, industry context rules and ironic expression rules. S4260: Based on the semantic collaboration results, or based on the semantic conflict results and weight values, determine the true intention of the community opinion leaders' statements, and determine the core viewpoints, sentiment tags, and potential market influence tags based on the true intentions.

[0058] Simple keyword matching is insufficient to handle irony, metaphor, or multimodal expressions in financial communities. Therefore, it is necessary to determine the true intentions of community opinion leaders' statements through semantic signal extraction, entity recognition, consistency judgment, and conflict resolution.

[0059] In practical implementation, semantic signals can be extracted through word segmentation, emoticon recognition, code keyword matching, and industry jargon dictionary extraction. Technical terms and non-standard abbreviations in plain text can be identified through dictionaries or semantic models; emoticons can be identified as corresponding emotional signals, such as a rocket emoticon indicating bullish sentiment, and a crying emoticon indicating loss or sarcasm; code snippets can be identified through programming language keywords or function calls; and industry jargon can be mapped to standard financial semantics through a pre-set dictionary.

[0060] When identifying financial entity information in various semantic signals, specific token symbols, project names, or financial product categories can be identified; opinion information is used to express attitudes toward specific entities, sentiment information is used to express emotional intensity, and market impact information is used to express potential market behavior.

[0061] When calculating the correlation strength and consistency between different semantic signals, the system analyzes whether the financial entity information pointed to by each semantic signal is consistent, whether the opinion information is in the same direction, whether the sentiment information matches, and whether the market impact information supports each other. The correlation strength can be determined based on the distance and co-occurrence frequency of the semantic signals in the text; consistency is used to determine whether different semantic signals form the same semantic direction. If the opinion of a text fragment is "bullish" and the emoticon is a "rocket emoticon" or a "diamond hand emoticon", then a semantic synergy result can be obtained; if the text fragment is "to the moon" and a "crying emoticon" appears at the same time, then a semantic conflict result can be obtained.

[0062] The pre-defined contextual knowledge base may include financial community expression patterns, industry context rules, and ironic expression rules, which are used to determine the weight of different semantic signals when semantic signals conflict. For example, when bullish text is paired with a negative emoji, the ironic expression rule can increase the weight of the semantic signal corresponding to the emoji and decrease the weight of the semantic signal corresponding to the text fragment.

[0063] When a semantic conflict is detected, the system determines the text weight and emoji weight based on a pre-defined contextual knowledge base, and calculates the sentiment polarity: SentPol = Wtext × BullScore + Wemoji × BearScore, where SentPol is the sentiment polarity, Wtext is the text weight, Wemoji is the emoji weight, BullScore is the positive bullish score corresponding to the text segment, and BearScore is the negative bearish score corresponding to the emoji. BullScore is greater than 0, BearScore is less than 0, Wtext and Wemoji are both non-negative, and Wtext + Wemoji = 1. A SentPol greater than 0 indicates a bullish bias, while a SentPol less than 0 indicates a bearish or sarcastic bias. Based on the semantic collaboration results, or based on the semantic conflict results and weight values, the system determines the true intention of the community opinion leader's statements and further identifies the core viewpoint, sentiment label, and potential market influence label.

[0064] Preferably, the preset scenario knowledge base is updated, including: S4251: Continuously acquire usage data, which includes community comments containing internet slang and emojis, the posting time of the community comments, the posting user information, and the context of the community comments. S4252: Based on usage data, statistically analyze the frequency of occurrence, co-occurring words, and corresponding emotional tendencies of internet slang and emoticons in different contexts, and identify the changes in the meaning of internet slang and emoticons based on the statistical results; S4253: When a change in meaning of internet slang or emoticons is detected, collect contextual corpus containing internet slang or emoticons and obtain interactive feedback from community users on the contextual corpus. The interactive feedback includes at least one of comments, explanations, clarifications, corrections, and forwarding evaluations. S4254: Based on the contextual corpus and interactive feedback, update at least one of the following in the preset scenario knowledge base: financial community expression patterns, industry context rules, and ironic expression rules corresponding to internet slang or emoticons; S4255: When evaluating the weights of conflicting semantic signals based on a preset contextual knowledge base, at least one of the updated financial community expression patterns, industry context rules, and ironic expression rules is invoked to determine the weight values ​​corresponding to the conflicting semantic signals, and the true intentions of the community opinion leaders' statements are determined based on the weight values.

[0065] Online language and community expression habits are constantly changing, and static pre-set contextual knowledge bases are difficult to adapt to new expression patterns. Therefore, it is necessary to update the pre-set contextual knowledge bases based on usage data.

[0066] In its implementation, the system continuously acquires community posts containing internet slang and emoticons, and divides them into sliding time windows based on the posting time. It also distinguishes different user groups or user sources based on the posting user information. The system analyzes the frequency of internet slang and emoticons in different contexts, co-occurring words, and corresponding emotional tendencies to identify changes in the meaning of internet slang or emoticons across different time windows or user groups.

[0067] Meaning change detection can be achieved by calculating the KL divergence of co-occurring word distributions within adjacent sliding time windows. The calculation formula is: KLD = Σ_i p_i×ln(p_i / q_i), where KLD is the KL divergence value, p_i is the probability of the i-th co-occurring word in the current sliding time window, q_i is the probability of the i-th co-occurring word in the previous sliding time window, and i is the co-occurring word index. When q_i is 0, a preset smoothing factor ε can be used to smooth p_i and q_i to avoid the formula failing to calculate. The larger the KLD, the more obvious the change in the usage context of internet slang or emoticons.

[0068] When a change in meaning is detected, the system collects contextual data containing the internet slang or emoji, and obtains interactive feedback from community users, including comments, explanations, clarifications, corrections, and reposts. The system can identify the intent to clarify or correct based on keywords, sentiment, and the authority of the user providing the feedback. The authority of the user can be determined based on whether the user is a verified account, their historical accuracy, or the quality of their past interactions.

[0069] Based on contextual corpora and interactive feedback, the system updates the corresponding financial community expression patterns, industry context rules, or ironic expression rules in the preset scenario knowledge base. For example, when a certain emoji changes from a bullish meaning to a warning meaning in a new context, the rule item corresponding to that emoji can be updated, and its weight value in semantic conflict handling can be adjusted.

[0070] In subsequent semantic conflict handling, the system calls the updated financial community expression mode, industry context rules or ironic expression rules to evaluate the weight of conflicting semantic signals, and determines the true intention of the community opinion leader's statement based on the weight value, thereby reducing misjudgment caused by the lag in the meaning of online slang or emoticons.

[0071] Preferably, based on the fields for the affected object, the direction of risk impact, the direction of attractiveness impact, the intensity of risk impact, and the intensity of attractiveness impact in the market event record, the adjustment amounts for the risk level and attractiveness corresponding to the affected object are determined, including: S5100: Determine the target financial product based on the related financial product or related product category; S5200: Obtain product operation status data of the target financial product within a preset time window. The product operation status data includes current risk level data, market liquidity data, and volatility data. Among them, the current risk level data includes at least one of the target financial product's current risk level, current risk coefficient, compliance risk label, and credit risk score. The market liquidity data includes at least one of the target financial product's trading volume, turnover, turnover rate, bid-ask spread, subscription and redemption success rate, and liquidity period. The volatility data includes at least one of the target financial product's price volatility, return standard deviation, price amplitude, and maximum drawdown. S5300: Normalize and weight the operational status data of each product, and calculate the context sensitivity factor of the target financial product. S5400: Determine the risk adjustment direction corresponding to the target financial product based on the risk impact direction field in the market event record; S5500: Determine the direction of attractiveness adjustment for the target financial product based on the attractiveness influence direction field in the market event record; S5600: Determine the risk level adjustment amount for the target financial product based on the risk adjustment direction, risk impact intensity field, and context sensitivity factor; S5700: Determine the attractiveness adjustment amount of the target financial product based on the attractiveness adjustment direction, the attractiveness influence strength field, and the context sensitivity factor.

[0072] Simply relying on community sentiment analysis to adjust product parameters may be out of touch with the actual operating status of the product. Therefore, it is necessary to combine various fields in the market event records with the product operating status data of the target financial product to jointly determine the adjustment amount for risk level and attractiveness.

[0073] In practice, the "Affected Object" field in the market event record can be used to characterize related financial products or related product categories. When the "Affected Object" field corresponds to a specific related financial product, that related financial product is identified as the target financial product; when the "Affected Object" field corresponds to a related product category, the financial products under that category are identified as the target financial product. Therefore, the target financial product originates from the "Affected Object" field in the market event record.

[0074] After identifying the target financial product, the system acquires its operational status data within a preset time window. This preset time window can be set according to the product's volatility characteristics; for example, a shorter time window is used for high-volatility assets, and a longer time window for stable assets. The product operational status data includes current risk level data, market liquidity data, and volatility data. The current risk level data characterizes the underlying risk status of the target financial product, the market liquidity data characterizes its tradability or realizability, and the volatility data characterizes the degree of price or return volatility of the target financial product.

[0075] When normalizing product operation status data, each indicator can be mapped to the range of 0 to 1. For current risk level data, the normalized value increases with higher risk; for market liquidity data, the market liquidity risk normalized value is used, with a larger value indicating poorer liquidity; for volatility data, the normalized value increases with higher volatility. Trading volume can be standardized on a logarithmic scale: NormTradeVol = log(TradeVol+1) / log(MaxTradeVol+1), where NormTradeVol is the standardized trading volume value, TradeVol is the trading volume of the target financial product, and MaxTradeVol is the historical maximum trading volume of similar products.

[0076] When calculating the weighted situation sensitivity factor, the formula can be used: F = Wrisk×RiskComp + Wliq×LiqRiskComp + Wvol×VolComp, where F is the situation sensitivity factor, RiskComp is the comprehensive risk level value, LiqRiskComp is the comprehensive market liquidity risk value, VolComp is the comprehensive volatility value, Wrisk is the weight of the comprehensive risk level value, Wliq is the weight of the comprehensive market liquidity risk value, Wvol is the weight of the comprehensive volatility value, Wrisk, Wliq and Wvol are all non-negative numbers, and Wrisk+Wliq+Wvol=1.

[0077] RiskComp can be obtained by weighting the normalized results of the current risk level, current risk coefficient, compliance risk label, and credit risk score; LiqRiskComp can be obtained by weighting the normalized results of trading volume, turnover, turnover rate, bid-ask spread, subscription and redemption success rate, and realizable period; VolComp can be obtained by weighting the normalized results of price volatility, standard deviation of return, price amplitude, and maximum drawdown.

[0078] The risk adjustment direction is determined based on the "Risk Impact Direction" field in the market event record: when the "Risk Impact Direction" field indicates increased risk, the risk adjustment direction is increased; when the "Risk Impact Direction" field indicates decreased risk, the risk adjustment direction is decreased; and when the "Risk Impact Direction" field indicates maintained risk, the risk adjustment direction is maintained. The attractiveness adjustment direction is determined based on the "Attractiveness Impact Direction" field in the market event record: when the "Attractiveness Impact Direction" field indicates enhanced attractiveness, the attractiveness adjustment direction is increased; when the "Attractiveness Impact Direction" field indicates weakened attractiveness, the attractiveness adjustment direction is decreased; and when the "Attractiveness Impact Direction" field indicates maintained attractiveness, the attractiveness adjustment direction is maintained.

[0079] When determining the risk level adjustment amount, the formula can be used: ΔRisk = kRisk × Ir × F, where ΔRisk is the risk level adjustment amount, kRisk is the risk adjustment direction coefficient, Ir is the risk impact intensity value corresponding to the risk impact intensity field, and F is the situation sensitivity factor. When the risk adjustment direction is increasing, kRisk is +1; when the risk adjustment direction is decreasing, kRisk is -1; and when the risk adjustment direction is maintaining, kRisk is 0. Therefore, the risk level adjustment amount is simultaneously affected by the risk impact direction field, the risk impact intensity field, and the situation sensitivity factor.

[0080] When determining the attraction adjustment amount, the formula can be used: ΔAttr = kAttr × Ia × F, where ΔAttr is the attraction adjustment amount, kAttr is the attraction adjustment direction coefficient, Ia is the attraction influence strength value corresponding to the attraction influence strength field, and F is the situation sensitivity factor. When the attraction adjustment direction is increasing, kAttr is +1; when the attraction adjustment direction is decreasing, kAttr is -1; and when the attraction adjustment direction is maintaining, kAttr is 0. Therefore, the attraction adjustment amount is simultaneously affected by the attraction influence direction field, the attraction influence strength field, and the situation sensitivity factor.

[0081] This application further proposes a method for calculating context sensitivity based on product type differentiation weights. In practical applications, this method calculates the context sensitivity factor of the target financial product by normalizing and weighting the operational status data of each product. The specific implementation process is as follows: S5310: Obtain the product type identifier of the target financial product. The product type identifier is used to indicate the type of financial product to which the target financial product belongs.

[0082] S5320: Based on the product type identifier, retrieve the risk level weight, market liquidity weight, and volatility weight corresponding to the financial product type from the preset product type weight parameter table.

[0083] S5330: Normalize the current risk level data to obtain the normalized risk level value.

[0084] S5340: Normalize market liquidity data to obtain normalized market liquidity values.

[0085] S5350: Normalizes volatility data to obtain normalized volatility values.

[0086] S5360: Multiply the normalized risk level value with the risk level weight to obtain the result of the first multiplication operation.

[0087] S5370: Multiply the normalized market liquidity value by the market liquidity weight to obtain the result of the second multiplication operation.

[0088] S5380: Multiply the volatility normalized value with the volatility weight to obtain the result of the third multiplication operation.

[0089] S5390: Sum the results of the first, second, and third multiplication operations to obtain the context sensitivity factor of the target financial product.

[0090] In practical applications, product type identifiers can use a hierarchical coding system to represent financial product types. For example, CRYPTO_DEFI represents decentralized finance (DeFi) crypto assets, EQUITY_GROWTH represents growth stock funds, BOND_GOVT represents government bonds, and MONEY_MARKET represents money market funds. These product type identifiers can be entered into the basic information table of the financial product database and used to retrieve the corresponding risk level weight, market liquidity weight, and volatility weight.

[0091] A pre-defined product type weight parameter table stores the weight parameters corresponding to different financial product types. It can be constructed based on the correlation between risk levels, market liquidity, volatility, and actual risk events in the historical operating data of various financial products. An example of its structure is shown below: Table 1: Parameter Table CRYPTO_DEFI 0.30 0.20 0.50 2024-01-15 EQUITY_GROWTH 0.35 0.25 0.40 2024-01-15 BOND_GOVT 0.40 0.45 0.15 2024-01-15 MONEY_MARKET 0.30 0.60 0.10 2024-01-15 When normalizing the current risk level data, for risk grade data, the mapping formula can be used: RiskGradeNorm = 0.2 + (RiskLevelNum-1)×0.2, where RiskGradeNorm is the normalized risk grade value, RiskLevelNum is the risk grade value, and R1 to R5 correspond to 1 to 5 respectively. For continuous variables such as risk coefficients, standardization can be performed using Zrisk = (RiskRaw-μrisk) / σrisk, where Zrisk is the standardized risk coefficient value, RiskRaw is the original risk coefficient value, μrisk is the historical risk coefficient sample mean of similar products, and σrisk is the historical risk coefficient sample standard deviation of similar products. Zrisk is then mapped to the 0 to 1 interval through a preset mapping function or pruning method.

[0092] When normalizing market liquidity data, to ensure that a higher scenario sensitivity factor value indicates greater sensitivity of the target financial product to market sentiment shocks, the market liquidity normalization value is calculated in the direction of liquidity risk; that is, a higher value indicates worse liquidity. For indicators such as trading volume, turnover, turnover rate, and subscription / redemption success rate, which indicate better liquidity, they can be standardized first and then subtracted from the standardized value by 1. For indicators such as bid-ask spread and liquidity period, which indicate worse liquidity, they can be directly calculated using LiqNorm = (LiqRaw - LiqMin) / (LiqMax - LiqMin), where LiqNorm is the market liquidity normalization value, LiqRaw is the original value of the market liquidity indicator, and LiqMin and LiqMax are the minimum and maximum values ​​of the corresponding indicators for similar products within a preset historical range, respectively.

[0093] When normalizing volatility data, it can be calculated as VolNorm = (VolRaw-VolMin) / (VolMax-VolMin), where VolNorm is the normalized volatility value, VolRaw is the original value of the volatility index, and VolMin and VolMax are the minimum and maximum values ​​of the corresponding volatility index for the same product within a preset historical range, respectively. To reduce the impact of extreme outliers, the 5th and 95th percentiles of historical volatility can also be used as VolMin and VolMax, respectively.

[0094] Subsequently, the normalized risk level, normalized market liquidity, and normalized volatility values ​​are multiplied by their respective weights, and the results of these three multiplications are summed to obtain the situation sensitivity factor. Its calculation formula is: F = RiskNorm × Wrisk + LiqNorm × Wliq + VolNorm × Wvol, where F is the situation sensitivity factor, RiskNorm is the normalized risk level value, LiqNorm is the normalized market liquidity value, VolNorm is the normalized volatility value, Wrisk is the risk level weight, Wliq is the market liquidity weight, and Wvol is the volatility weight. Wrisk, Wliq, and Wvol are all non-negative, and Wrisk + Wliq + Wvol = 1.

[0095] For example, if a decentralized finance token has a RiskNorm of 0.8, a LiqNorm of 0.7, and a VolNorm of 0.9, and corresponding Wrisk of 0.30, Wliq of 0.20, and Wvol of 0.50, then F = 0.8 × 0.30 + 0.7 × 0.20 + 0.9 × 0.50 = 0.83, indicating that the target financial product has high context sensitivity.

[0096] Furthermore, based on the product type identifier, after retrieving the risk level weight, market liquidity weight, and volatility weight corresponding to the financial product type from the preset product type weight parameter table, it also includes: S5321: Obtain the current market cycle indicator, which includes bull market indicator, bear market indicator, or sideways market indicator; S5322: Based on the current market cycle identifier, retrieve the risk level correction coefficient, market liquidity correction coefficient, and volatility correction coefficient from the preset market cycle correction parameter table; S5323: Adjust the risk level weights according to the risk level correction coefficient to obtain the target risk level weights; S5324: Adjust the market liquidity weights according to the market liquidity adjustment coefficient to obtain the target market liquidity weights; S5325: Adjust the volatility weights according to the volatility correction coefficient to obtain the target volatility weights; S5326: Use the target risk level weight, target market liquidity weight, and target volatility weight to calculate the context sensitivity factor of the target financial product.

[0097] In practice, the method for obtaining the current market cycle identifier will be detailed later. A pre-defined market cycle correction parameter table stores the risk level correction coefficient, market liquidity correction coefficient, and volatility correction coefficient for different market cycles. For example, under the market cycle corresponding to a bull market identifier, the risk level correction coefficient can be reduced while the market liquidity correction coefficient can be increased; under the market cycle corresponding to a bear market identifier, both the risk level correction coefficient and the volatility correction coefficient can be increased; and under the market cycle corresponding to a sideways market identifier, all correction coefficients can be kept at 1.

[0098] The correction calculation can employ multiplicative correction and normalization. Let Wrisk be the risk level weight, Wliq be the market liquidity weight, Wvol be the volatility weight, CorrRisk be the risk level correction coefficient, CorrLiq be the market liquidity correction coefficient, CorrVol be the volatility correction coefficient, and WsumCycle be the sum of the weights after market cycle correction. Then, WsumCycle = Wrisk × CorrRisk + Wliq × CorrLiq + Wvol × CorrVol; the target risk level weight WriskTarget = (Wrisk × CorrRisk) / WsumCycle, the target market liquidity weight WliqTarget = (Wliq × CorrLiq) / WsumCycle, and the target volatility weight WvolTarget = (Wvol × CorrVol) / WsumCycle. Here, CorrRisk, CorrLiq, and CorrVol are all greater than 0 to ensure that WsumCycle is greater than 0.

[0099] When calculating the context sensitivity factor for the target financial product, the risk level weight, market liquidity weight, and volatility weight are replaced with the target risk level weight, target market liquidity weight, and target volatility weight, respectively. The corresponding calculation formula is: F = RiskNorm×WriskTarget + LiqNorm×WliqTarget + VolNorm×WvolTarget, where F is the context sensitivity factor, RiskNorm is the normalized risk level value, LiqNorm is the normalized market liquidity value, VolNorm is the normalized volatility value, WriskTarget is the target risk level weight, WliqTarget is the target market liquidity weight, and WvolTarget is the target volatility weight.

[0100] In a preferred embodiment of this application, obtaining the current market cycle identifier includes: S53211: Obtain emerging financial market data, which includes at least one of the following: number of on-chain transactions of digital assets, number of active addresses, and trading depth of a specific digital asset. S53212: Obtain traditional financial market indicators, including at least one of the following: daily changes in major stock indices, trends in the bond yield curve, and price movements of commodities. S53213: Perform nonlinear trend analysis on emerging financial market data to obtain emerging market trend analysis results used to represent potential market inflection points or abnormal fluctuation patterns. S53214: Perform trend and volatility analysis on traditional financial market indicators to obtain traditional market analysis results used to represent macro market sentiment or capital flows; S53215: Integrate the results of emerging market trend analysis with the results of traditional market analysis to obtain market cycle determination data; S53216: Determine the current market cycle identifier based on market cycle judgment data and preset market cycle division rules.

[0101] In practice, emerging financial market data can be obtained through blockchain data interfaces, on-chain data analysis platforms, or transaction data interfaces. The number of on-chain transactions reflects network activity, the number of active addresses reflects user participation, and the trading depth of a specific digital asset reflects market capacity; the trading depth can be the sum of buy and sell orders.

[0102] Nonlinear trend analysis can employ the Hurst index method to identify trend persistence in emerging financial market data. To avoid confusion with other symbols throughout the text, the Hurst index is denoted as HurstVal. For the emerging financial market data time series ChainSeq(t), the rescaled range RangeH(n) / StdH(n) at different time scales n is calculated, and HurstVal is obtained through double log-log regression. Here, ChainSeq(t) is the emerging financial market data series at time t, n is the time scale, RangeH(n) is the range at that time scale, and StdH(n) is the standard deviation at that time scale. A HurstVal greater than 0.5 indicates trend persistence, a HurstVal less than 0.5 indicates mean regression, and a HurstVal equal to 0.5 indicates a random walk. When StdH(n) is 0, that time scale can be skipped or a preset smoothing factor εH can be used to avoid the rescaled range being uncalculated.

[0103] A HurstVal greater than 0.5 only indicates trend persistence. Whether it belongs to a bull or bear market trend needs to be determined in conjunction with the number of on-chain transactions, the number of active addresses, transaction depth, and the direction of price series changes. The system can also use a change point detection algorithm to identify trend turning points, so as to obtain emerging market trend analysis results that can be used to represent potential market turning points or abnormal fluctuation patterns.

[0104] Traditional financial market indicators can include major stock indices, bond yield curves, and commodity price trends. Trend analysis can use the Moving Average Convergence Divergence (MACD) indicator, where EMA12 is the 12-day exponential moving average, EMA26 is the 26-day exponential moving average, DIFmacd = EMA12 - EMA26, and DEAmacd is the 9-day exponential moving average of DIFmacd. A crossover of DIFmacd above DEAmacd indicates a bullish signal, while a crossover of DIFmacd below DEAmacd indicates a bearish signal.

[0105] Volatility analysis can use the Average True Range (ATR), where ATRtrad is the average true range corresponding to traditional financial market indicators, TRtrad is the true range, and TRtrad is the maximum value among the differences between the highest and lowest prices of the day, the differences between the previous day's closing price and the highest price of the day, and the differences between the previous day's closing price and the lowest price of the day. ATRtrad is the moving average of TRtrad over a preset number of days, used to represent the volatility intensity of traditional financial market indicators.

[0106] The fusion process can be implemented using a weighted voting mechanism or a classification model. For example, when both emerging market trend analysis and traditional market analysis results show an upward trend and volatility is within a preset moderate range, a bull market is identified; when both show a downward trend, and HurstVal indicates the downward trend is persistent, a bear market is identified; when the signals from both are contradictory or neither shows a clear trend, a sideways market is identified. The resulting market cycle determination data can include emerging market trend analysis results, traditional market analysis results, volatility quantile results, and the consistency judgment results of the two, and serve as input to the preset market cycle division rules.

[0107] Furthermore, nonlinear trend analysis is performed on emerging financial market data to obtain emerging market trend analysis results used to represent potential market inflection points or abnormal fluctuation patterns, including: S532131: Real-time collection and cleaning of emerging financial market data to obtain cleaned emerging financial market data; S532132: Based on cleaned emerging financial market data, nonlinear trend analysis is performed to obtain emerging market trend analysis results that can be used to represent potential market inflection points or abnormal fluctuation patterns. S532141: Standardize traditional financial market indicators to obtain standardized traditional financial market indicators. S532142: Based on standardized traditional financial market indicators, trend analysis and volatility analysis are performed to obtain traditional market analysis results used to represent macro market sentiment or capital flows; S53215 includes: S532151: Determine the first fusion weight based on the real-time update frequency and data quality assessment results of the cleaned emerging financial market data; S532152: Determine the second fusion weight based on the update frequency and data quality assessment results of standardized traditional financial market indicators; S532153: The results of emerging market trend analysis and traditional market analysis are weighted and merged according to the first fusion weight and the second fusion weight to obtain fused data; S532154: Perform consistency verification on the fused data and obtain the verification result; S532155: When the verification result meets the preset consistency conditions, the fused data will be determined as market cycle determination data; S532156: When the verification result does not meet the preset consistency conditions, abnormal data items in the fused data are removed or downweighted, and the processed fused data is determined as market cycle judgment data.

[0108] In practice, data cleaning can remove abnormally active addresses, abnormal orders that deviate significantly from market prices, and other data items that do not meet preset data quality conditions based on the characteristics of on-chain data. The cleaned data can then be scored based on the missing rate, outlier ratio, and data latency. The calculation formula is: QualScore = 0.4×(1-MissRate) + 0.4×(1-AbnRate) + 0.2×max(0, 1-DelayTime / DelayMax), where QualScore is the data quality score, MissRate is the missing rate, AbnRate is the outlier ratio, DelayTime is the data latency, and DelayMax is the maximum tolerable latency (DelayMax is greater than 0). MissRate and AbnRate both range from 0 to 1, and max(0, 1-DelayTime / DelayMax) is used to avoid negative values ​​when the data latency exceeds the maximum tolerable latency.

[0109] The first and second fusion weights can be determined based on data quality scores and update timeliness. The real-time update frequency can be represented by the update interval in minutes; the shorter the update interval, the higher the timeliness. Specifically, the unnormalized fusion weights can be calculated first: WnewRaw = 0.6×QualNew + 0.4×min(1, 1 / ln(UpdateNew+1)), WtradRaw =0.6×QualTrad + 0.4×min(1, 1 / ln(UpdateTrad+1)), where WnewRaw is the first unnormalized fusion weight, WtradRaw is the second unnormalized fusion weight, QualNew is the data quality score of the cleaned emerging financial market data, QualTrad is the data quality score of the standardized traditional financial market indicator, UpdateNew is the update interval in minutes for the emerging financial market data, UpdateTrad is the update interval in minutes for the traditional financial market indicator, and both UpdateNew and UpdateTrad are greater than 0. Then, normalization yields Wnew = WnewRaw / (WnewRaw+WtradRaw), Wtrad = WtradRaw / (WnewRaw+WtradRaw), where Wnew is the first fusion weight, Wtrad is the second fusion weight, and Wnew+Wtrad=1.

[0110] When weighting the data according to the first and second fusion weights, the formula can be used: CycleFusion = Wnew × TrendNew + Wtrad × TrendTrad, where CycleFusion is the fused data, TrendNew is the numerical data corresponding to the emerging market trend analysis results, TrendTrad is the numerical data corresponding to the traditional market analysis results, Wnew is the first fusion weight, and Wtrad is the second fusion weight. Therefore, data sources with higher data quality scores and shorter update intervals have higher weights in the fused data.

[0111] The consistency check calculates the correlation coefficient between the emerging market trend analysis results and the traditional market analysis results. The formula is: CorrTrend = Σ[(NewTrend_i-NewTrendAvg)×(TradTrend_i-TradTrendAvg)] / √[Σ(NewTrend_i-NewTrendAvg)²×Σ(TradTrend_i-TradTrendAvg)²], where CorrTrend is the correlation coefficient between the emerging market trend analysis results and the traditional market analysis results, NewTrend_i is the emerging market trend score at the i-th time point, TradTrend_i is the traditional market trend score at the i-th time point, NewTrendAvg is the average of the emerging market trend score sequence, TradTrendAvg is the average of the traditional market trend score sequence, and i is the time point number. When the denominator of the formula is 0, the check result is determined to not meet the preset consistency condition.

[0112] When CorrTrend is greater than 0.6, the verification result is determined to meet the preset consistency condition, and the fused data is identified as market cycle determination data; when CorrTrend is between 0.3 and 0.6, the emerging market trend analysis result is determined to be inconsistent with the traditional market analysis result, and data items with large deviations are downweighted; when CorrTrend is less than 0.3, the two are determined to be seriously divergent, and abnormal data items can be removed, the weight of abnormal data items can be reduced, or the result with the higher data quality score can be used as the processed fused data.

[0113] Reference Figure 2 This application also proposes a personalized financial service recommendation system based on a large model, including: Customer feature data acquisition module 1 is used to acquire customer feature data of target customers. The customer feature data includes structured financial behavior data and unstructured financial expression data of target customers. The initial recommendation set generation module 2 is used to input customer feature data into the large model recommendation model, generate customer profile feature vectors of target customers, and generate an initial recommendation set based on the matching relationship between the customer profile feature vectors and the product feature vectors of each financial product in the financial product library. The large model recommendation model is a model used for semantic understanding, behavioral analysis and preference extraction of customer feature data. The initial recommendation set includes at least one financial product and the original recommendation score corresponding to the financial product. External market information acquisition module 3 is used to acquire external market information through external data interfaces. External market information includes macroeconomic data, financial regulatory policy information, and breaking news events. Market event analysis module 4 is used to analyze external market information and generate at least one market event record. Each market event record includes fields for the affected object, the direction of risk impact, the direction of attractiveness impact, the intensity of risk impact, and the intensity of attractiveness impact. The market status parameter adjustment module 5 is used to determine the risk level adjustment amount and attractiveness adjustment amount corresponding to the affected object based on the affected object field, risk impact direction field, attractiveness impact direction field, risk impact intensity field and attractiveness impact intensity field in the market event record, and update the market status parameters corresponding to the affected object in the real-time market status parameter table according to the risk level adjustment amount and attractiveness adjustment amount. The real-time market status parameter table is used to store the market status parameters corresponding to financial products or financial product categories. The market status parameters include risk level and attractiveness. The calibration recommendation score calculation module 6 is used to read the corresponding risk level and attractiveness from the real-time market status parameter table for financial products in the initial recommendation set, and calculate the calibration recommendation score based on the original recommendation score, risk level and attractiveness. The recommendation set correction module 7 is used to correct the initial recommendation set based on the risk level, attractiveness, and calibration recommendation score to obtain the target recommendation set. The correction includes: filtering financial products with a risk level higher than a preset risk threshold, and / or filtering financial products with an attractiveness lower than a preset attractiveness threshold, and reordering the unfiltered financial products according to the calibration recommendation score.

[0114] In the aforementioned system, the customer feature data acquisition module 1, the initial recommendation set generation module 2, the external market information acquisition module 3, the market event analysis module 4, the market state parameter adjustment module 5, the calibration recommendation score calculation module 6, and the recommendation set correction module 7 respectively execute the customer feature data acquisition, initial recommendation set generation, external market information acquisition, market event record generation, market state parameter update, calibration recommendation score calculation, and target recommendation set generation processes in the aforementioned method. Through the coordinated operation of these modules, a complete processing chain can be formed, encompassing customer profile analysis, market event analysis, risk level and attractiveness updates, and recommendation set correction.

[0115] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A personalized financial service recommendation method based on a large model, characterized in that, include: Acquire customer characteristic data of target customers, including structured financial behavior data and unstructured financial expression data of target customers; The customer feature data is input into the large model recommendation model to generate a customer profile feature vector of the target customer. Based on the matching relationship between the customer profile feature vector and the product feature vector of each financial product in the financial product library, an initial recommendation set is generated. The large model recommendation model is a model used to perform semantic understanding, behavioral analysis and preference extraction on the customer feature data. The initial recommendation set includes at least one financial product and the original recommendation score corresponding to the financial product. External market information is obtained through external data interfaces, including macroeconomic data, financial regulatory policy information, and breaking news events. The external market information is analyzed to generate at least one market event record. Each market event record includes an affected object field, a risk impact direction field, an attractiveness impact direction field, a risk impact intensity field, and an attractiveness impact intensity field. Based on the fields of the affected object, the direction of risk impact, the direction of attractiveness impact, the intensity of risk impact, and the intensity of attractiveness impact in the market event record, the risk level adjustment and attractiveness adjustment corresponding to the affected object are determined. Based on the risk level adjustment and the attractiveness adjustment, the market state parameters corresponding to the affected object in the real-time market state parameter table are updated. The real-time market state parameter table is used to store the market state parameters corresponding to financial products or financial product categories. The market state parameters include risk level and attractiveness. For the financial products in the initial recommendation set, the corresponding risk level and attractiveness are read from the real-time market status parameter table, and a calibration recommendation score is calculated based on the original recommendation score, the risk level, and the attractiveness. Based on the risk level, the attractiveness, and the calibration recommendation score, the initial recommendation set is modified to obtain the target recommendation set. The modification includes: filtering financial products with a risk level higher than a preset risk threshold, and / or filtering financial products with an attractiveness lower than a preset attractiveness threshold, and reordering the unfiltered financial products according to the calibration recommendation score.

2. The personalized financial service recommendation method based on a large model according to claim 1, characterized in that, The step of parsing the external market information to generate at least one market event record includes: The opinions of community opinion leaders are obtained from pre-defined emerging financial information sources, including at least one of technology forums, developer communities, crypto asset-related mailing lists, and industry analyst information sources. The speech of the community opinion leaders is subjected to domain semantic analysis to identify technical terms, informal expressions and industry jargon involved in the speech, and the core viewpoints, sentiment tags, potential market impact tags and related financial products or related product categories are determined according to the identification results. Obtain the influence score corresponding to the community opinion leader. The influence score is determined based on at least one of the following: the frequency of citation of the community opinion leader's remarks, the amount of content dissemination, the historical prediction accuracy, the amount of community interaction, and the activity level. Based on the core viewpoint, the sentiment tag, the potential market impact tag, the associated financial product or associated product category, and the influence score, a community viewpoint market event record is generated to form one of the market event records. The impact object field of the community viewpoint market event record is determined according to the associated financial product or associated product category. The risk impact direction field and the attractiveness impact direction field of the community viewpoint market event record are determined according to the sentiment tag and the potential market impact tag. The risk impact strength field and the attractiveness impact strength field of the community viewpoint market event record are determined according to the influence score.

3. The personalized financial service recommendation method based on a large model according to claim 2, characterized in that, The statements of the community opinion leaders are subjected to domain semantic analysis to identify technical terms, informal expressions, and industry jargon used in the statements. Based on the identification results, the core viewpoints, sentiment tags, potential market impact tags, and associated financial products or product categories are determined, including: Extract at least two types of semantic signals from the statements of the community opinion leaders, including technical terms, text fragments, emoticons, code snippets, non-standard abbreviations, or industry jargon; Identify information about financial entities, opinions, sentiments, and market impacts related to financial products or categories of financial products from various semantic signals. The financial entity information is matched with the financial product identifiers and / or financial product category identifiers in the financial product database to determine the associated financial product or associated product category corresponding to the statement. Based on the associated financial products or product categories, opinion information, sentiment information and market impact information corresponding to different semantic signals, the correlation strength and consistency between the different semantic signals are calculated to obtain semantic synergy results or semantic conflict results. When the semantic conflict result is obtained, the conflicting semantic signals are evaluated by weight according to the preset scenario knowledge base to obtain the weight value corresponding to the conflicting semantic signals. The preset scenario knowledge base includes at least one of financial community expression patterns, industry context rules and ironic expression rules. Based on the semantic collaboration results, or based on the semantic conflict results and the weight values, determine the true intention of the community opinion leader's statements, and determine the core viewpoint, the sentiment tag, and the potential market influence tag based on the true intention.

4. The personalized financial service recommendation method based on a large model according to claim 3, characterized in that, Updating the preset scenario knowledge base includes: Continuously acquire usage data, which includes community comments containing internet slang and emoticons, the posting time of the community comments, the posting user information, and the context content corresponding to the community comments; Based on the usage data, the frequency of occurrence, co-occurring words, and corresponding emotional tendencies of the internet slang and emoticons in different contexts are statistically analyzed, and the meaning changes of the internet slang and emoticons are identified based on the statistical results. When a change in meaning of the internet slang or emoji is detected, the contextual corpus containing the internet slang or emoji is collected, and interactive feedback generated by community users in response to the contextual corpus is obtained. The interactive feedback includes at least one of comments, explanations, clarifications, corrections, and forwarding evaluations. Based on the contextual corpus and the interactive feedback, update at least one of the financial community expression patterns, industry context rules, and ironic expression rules in the preset scenario knowledge base that correspond to the online slang or emoticons; When evaluating the weights of conflicting semantic signals based on the preset scenario knowledge base, at least one of the updated financial community expression patterns, industry context rules, and ironic expression rules is invoked to determine the weight values ​​corresponding to the conflicting semantic signals, and the true intentions of the community opinion leaders' statements are determined based on the weight values.

5. The personalized financial service recommendation method based on a large model according to claim 3, characterized in that, Based on the fields for the affected object, the direction of risk impact, the direction of attractiveness impact, the intensity of risk impact, and the intensity of attractiveness impact in the market event record, determine the risk level adjustment and attractiveness adjustment for the affected object, including: The target financial product is determined based on the aforementioned related financial product or related product category; The system acquires product operation status data of the target financial product within a preset time window. The product operation status data includes current risk level data, market liquidity data, and volatility data. The current risk level data includes at least one of the target financial product's current risk level, current risk coefficient, compliance risk indicator, and credit risk score. The market liquidity data includes at least one of the target financial product's trading volume, turnover, turnover rate, bid-ask spread, subscription and redemption success rate, and liquidation period. The volatility data includes at least one of the target financial product's price volatility, standard deviation of return, price amplitude, and maximum drawdown. The operational status data of each product are normalized and weighted to calculate the context sensitivity factor of the target financial product. Based on the risk impact direction field in the market event record, determine the risk adjustment direction corresponding to the target financial product; Based on the attraction influence direction field in the market event record, determine the attraction adjustment direction corresponding to the target financial product; The risk level adjustment amount of the target financial product is determined based on the risk adjustment direction, the risk impact intensity field, and the context sensitivity factor. The attractiveness adjustment amount of the target financial product is determined based on the attractiveness adjustment direction, the attractiveness influence strength field, and the context sensitivity factor.

6. The personalized financial service recommendation method based on a large model according to claim 5, characterized in that, The operational status data of each product are normalized and weighted to calculate the context sensitivity factor of the target financial product, including: Obtain the product type identifier of the target financial product, wherein the product type identifier is used to indicate the type of financial product to which the target financial product belongs; Based on the product type identifier, retrieve the risk level weight, market liquidity weight, and volatility weight corresponding to the financial product type from the preset product type weight parameter table; The current risk level data is normalized to obtain a normalized risk level value; The market liquidity data is normalized to obtain the normalized market liquidity value; The volatility data is normalized to obtain the volatility normalized value; The normalized value of the risk level is multiplied by the risk level weight to obtain the first multiplication result; The normalized market liquidity value is multiplied by the market liquidity weight to obtain the second multiplication result; The normalized volatility value is multiplied by the volatility weight to obtain the third multiplication result; The context sensitivity factor of the target financial product is obtained by summing the results of the first multiplication operation, the second multiplication operation, and the third multiplication operation.

7. The personalized financial service recommendation method based on a large model according to claim 6, characterized in that, After retrieving the risk level weight, market liquidity weight, and volatility weight corresponding to the financial product type from a preset product type weight parameter table based on the product type identifier, the process further includes: Obtain the current market cycle identifier, which includes a bull market identifier, a bear market identifier, or a sideways market identifier; Based on the current market cycle identifier, retrieve the risk level correction coefficient, market liquidity correction coefficient, and volatility correction coefficient from the preset market cycle correction parameter table; The risk level weights are adjusted according to the risk level correction coefficient to obtain the target risk level weights; The market liquidity weight is adjusted according to the market liquidity adjustment coefficient to obtain the target market liquidity weight. The volatility weight is adjusted according to the volatility correction coefficient to obtain the target volatility weight; The target risk level weight, the target market liquidity weight, and the target volatility weight are used to calculate the context sensitivity factor of the target financial product.

8. The personalized financial service recommendation method based on a large model according to claim 7, characterized in that, The step of obtaining the current market cycle identifier includes: Acquire emerging financial market data, which includes at least one of the following: the number of on-chain transactions of digital assets, the number of active addresses, and the trading depth of a specific digital asset; Obtain traditional financial market indicators, including at least one of the following: daily changes in major stock indices, trends in bond yield curves, and price movements of commodities. Nonlinear trend analysis is performed on the emerging financial market data to obtain emerging market trend analysis results that can be used to represent potential market inflection points or abnormal fluctuation patterns. Trend and volatility analyses are performed on the traditional financial market indicators to obtain traditional market analysis results used to represent macro market sentiment or capital flows. The emerging market trend analysis results are combined with the traditional market analysis results to obtain market cycle determination data; Based on the market cycle determination data and the preset market cycle division rules, the current market cycle identifier is determined.

9. The personalized financial service recommendation method based on a large model according to claim 8, characterized in that, Nonlinear trend analysis is performed on the emerging financial market data to obtain emerging market trend analysis results used to represent potential market inflection points or abnormal fluctuation patterns, including: The emerging financial market data is collected and cleaned in real time to obtain cleaned emerging financial market data. Based on the cleaned emerging financial market data, nonlinear trend analysis is performed to obtain emerging market trend analysis results that can represent potential market inflection points or abnormal fluctuation patterns. Trend and volatility analyses are performed on the aforementioned traditional financial market indicators to obtain traditional market analysis results used to represent macroeconomic market sentiment or capital flows, including: The traditional financial market indicators are standardized to obtain standardized traditional financial market indicators. Trend analysis and volatility analysis are then performed based on the standardized traditional financial market indicators to obtain traditional market analysis results that represent macro market sentiment or capital flows. The emerging market trend analysis results are integrated with the traditional market analysis results to obtain market cycle determination data, including: The first fusion weight is determined based on the real-time update frequency and data quality assessment results of the cleaned emerging financial market data; The second fusion weight is determined based on the update frequency and data quality assessment results of the standardized traditional financial market indicators. According to the first fusion weight and the second fusion weight, the emerging market trend analysis results and the traditional market analysis results are weighted and fused to obtain fused data; The fused data is subjected to a consistency check to obtain the check result; When the verification result meets the preset consistency condition, the fused data is determined as the market cycle determination data; When the verification result does not meet the preset consistency condition, abnormal data items in the fused data are removed or downweighted, and the processed fused data is determined as the market cycle determination data.

10. A personalized financial service recommendation system based on a large model, characterized in that, include: The customer feature data acquisition module is used to acquire customer feature data of target customers, including structured financial behavior data and unstructured financial expression data of target customers. The initial recommendation set generation module is used to input the customer feature data into the large model recommendation model, generate the customer profile feature vector of the target customer, and generate an initial recommendation set based on the matching relationship between the customer profile feature vector and the product feature vector of each financial product in the financial product library. The large model recommendation model is a model used to perform semantic understanding, behavioral analysis and preference extraction on the customer feature data. The initial recommendation set includes at least one financial product and the original recommendation score corresponding to the financial product. The external market information acquisition module is used to acquire external market information through an external data interface. The external market information includes macroeconomic data, financial regulatory policy information, and breaking news event information. The market event analysis module is used to analyze the external market information and generate at least one market event record. Each market event record includes an impact object field, a risk impact direction field, an attractiveness impact direction field, a risk impact strength field, and an attractiveness impact strength field. The market status parameter adjustment module is used to determine the risk level adjustment amount and attractiveness adjustment amount corresponding to the affected object based on the affected object field, risk impact direction field, attractiveness impact direction field, risk impact intensity field, and attractiveness impact intensity field in the market event record, and update the market status parameters corresponding to the affected object in the real-time market status parameter table based on the risk level adjustment amount and the attractiveness adjustment amount. The real-time market status parameter table is used to store the market status parameters corresponding to financial products or financial product categories, and the market status parameters include risk level and attractiveness. The calibration recommendation score calculation module is used to read the corresponding risk level and attractiveness from the real-time market status parameter table for the financial products in the initial recommendation set, and calculate the calibration recommendation score based on the original recommendation score, the risk level and the attractiveness. The recommendation set correction module is used to correct the initial recommendation set according to the risk level, the attractiveness, and the calibration recommendation score to obtain a target recommendation set. The correction includes: filtering financial products with a risk level higher than a preset risk threshold, and / or filtering financial products with an attractiveness lower than a preset attractiveness threshold, and reordering the unfiltered financial products according to the calibration recommendation score.

Images