Financial services product marketing method based on deepseek

Abstract

The present application relates to the technical field of intelligent marketing, in particular to a financial service product marketing method based on DeepSeek, comprising the following steps: collecting browsing time length and subscription interval sequence, calculating attention change to generate fluctuation index, combining policy hot words and yield rate data to compare deviation degree of prediction and real-time curve, using reinforcement learning to optimize recommendation ranking according to deviation sign and fluctuation interval, matching personalized combination of user and product feature vector, monitoring operation behavior to update time length sequence and analyzing behavior score. In the present application, the financial behavior conversion fluctuation index is tracked through a dynamic window, the market deviation degree of hot words and yield rate is combined, the recommendation ranking is adjusted according to the deviation state by reinforcement learning, precise recommendation is realized by feature vector matching, the strategy is continuously optimized through behavior monitoring closed-loop feedback, the intelligent decision chain from data collection to strategy adjustment is constructed, the limitations of traditional analysis and recommendation are broken through, and the marketing timeliness, precision and adaptability are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent marketing technology, and in particular to a marketing method for financial service products based on DeepSeek. Background Technology

[0002] The field of intelligent marketing technology encompasses technologies that utilize artificial intelligence, big data analytics, machine learning, and other information processing methods to conduct automated, personalized, and precise marketing decisions and execution based on user behavior data, transaction data, and external environmental variables. Core components of this technology include user profiling, behavioral prediction modeling, marketing strategy generation, marketing content delivery, and channel optimization. It emphasizes using technological means to refine user identification and classification, thereby implementing differentiated marketing. The systematic technical framework of this field involves data collection and processing, model building and training, strategy decision generation, information delivery execution, and effect tracking and feedback analysis, making it a crucial component closely integrated with marketing promotion in digital enterprise operations.

[0003] Financial service product marketing methods refer to approaches that target financial product users by mining and analyzing their financial attribute behavioral data to achieve product marketing strategy recommendations, target customer identification, and targeted content generation. The technical aspects covered in this topic include building correlation models to identify potential customers based on users' original product choices, account behavior, and risk preferences in financial scenarios; generating marketing strategy combinations based on the matching relationship between users and products; and dynamically updating user tags and recommended content in conjunction with current market data. The entire process is completed through steps such as modeling analysis, tag matching, strategy combination generation, and content synthesis.

[0004] Existing technologies rely on static analysis of raw behavioral data, lacking mechanisms to capture short-term fluctuations in user interests. This results in marketing strategies failing to respond promptly to changes in user needs. User profiling focuses on long-term behavioral patterns, ignoring external factors such as policy adjustments and competitor dynamics, leading to a disconnect between recommendation strategies and the market environment. Traditional recommendation systems use fixed-weight rules for matching, and strategy adjustments rely on manual intervention with long cycles, making them ill-suited to the high-frequency fluctuations of the financial market. Customer identification models are mostly based on single-dimensional data correlation analysis, failing to establish a multi-dimensional mapping relationship between user attributes and product characteristics, thus limiting the accuracy of personalized recommendations. The performance tracking stage suffers from data feedback lag, with behavioral data update frequency mismatched with strategy iteration cycles, creating data gaps in the decision-making chain. This leads to traditional marketing systems exhibiting poor recommendation timeliness, rigid strategies, and weak environmental adaptability in dynamic market environments. For example, when the central bank announces interest rate cuts, existing systems cannot quickly identify changes in user demand for bond products, continuing to recommend money market funds according to the established strategy, resulting in customer churn and decreased conversion rates. Summary of the Invention

[0005] To address the shortcomings of existing technologies that rely on static analysis of raw behavioral data and lack mechanisms to capture short-term fluctuations in user interests, resulting in marketing strategies failing to respond promptly to changes in user needs, and the tendency for user profiling to focus on long-term behavioral patterns while ignoring external factors such as policy adjustments and market competitor dynamics, leading to a disconnect between recommendation strategies and the market environment, this invention provides a financial service product marketing method based on DeepSeek. The technical solution is as follows: Traditional recommendation systems use fixed-weight rule matching, and strategy adjustments rely on manual intervention and have long cycles, making them ill-suited to the high-frequency fluctuations of the financial market. Customer identification models are often based on single-dimensional data correlation analysis, failing to establish a multi-dimensional mapping relationship between user attributes and product features, thus limiting the accuracy of personalized recommendations. Furthermore, the performance tracking stage suffers from data feedback lag, with the frequency of behavioral data updates mismatched with the strategy iteration cycle, creating data gaps in the decision-making chain. This results in traditional marketing systems exhibiting poor recommendation timeliness, rigid strategies, and weak environmental adaptability in dynamic market environments. For example, when a central bank interest rate cut is announced, existing systems cannot quickly identify changes in user demand for bond products and continue to recommend money market funds according to the established strategy, leading to customer churn and decreased conversion rates. This invention provides a financial service product marketing method based on DeepSeek. The technical solution is as follows:

[0006] On the one hand, it provides a marketing method for financial service products based on DeepSeek, which includes:

[0007] S1: By collecting the user's financial product browsing time sequence and fund subscription interval sequence, the dynamic sliding window algorithm based on the user behavior decay factor is called to calculate the change in product attention over three consecutive natural days, and generate a financial behavior volatility index.

[0008] S2: By obtaining the frequency of hot words in the central bank's policy texts and the change in yields of peer competitors, the dynamic weighting method is used to compare the demand forecast curve of bond products with the real-time subscription volume curve to obtain the deviation of market expectations.

[0009] S3: Based on the positive and negative signs of the market expectation deviation and the range of financial behavior volatility index values, the recommendation priority of bank wealth management products is reordered using the near-end strategy optimization method to generate a wealth management product recommendation queue.

[0010] S4: Based on the sorting results of the financial product recommendation queue, the similarity between the user profile vector and the feature vector of the recommended products is calculated using a feature vector matching algorithm, and a customized financial product combination is output.

[0011] S5: By monitoring the user's operational behavior regarding the customized financial product portfolio, update the product browsing time sequence, extract user behavior characteristics, and combine time series analysis methods to obtain a score for changes in user financial behavior.

[0012] As a further aspect of the present invention, the dynamic sliding window algorithm establishes a parameter correlation between the user behavior decay factor and the sliding window size, so that the weight of the behavior data at different time points within the window decreases over time, dynamically adjusts the sensitivity of user behavior in the time dimension, and finely controls the calculation accuracy of the change in attention.

[0013] The time series analysis method uses an ARIMA model with time-varying parameters, and its parameter update period is set to T+1 trading days, where T is the time interval between the current trading day and the most recent model training day.

[0014] The financial behavior volatility index includes changes in the number of product categories monitored, fluctuations in browsing frequency, and trends in behavioral activity. The market expectation deviation specifically includes the range of subscription volume exceeding expectations, the range of subscription volume falling short of expectations, and the range of subscription volume meeting expectations. The wealth management product recommendation queue includes priority product identifiers, product ranking weight values, and recommendation timeliness parameters. The customized financial product portfolio specifically refers to the set of products with the best user preference matching, the risk-return balance index among products, and the validity period setting of the customized portfolio. The user financial behavior change score includes browsing behavior stage change characteristics, subscription behavior stage transition signals, and behavior stability assessment indicators.

[0015] As a further aspect of the present invention, the specific steps of S1 include:

[0016] S101: Collect user browsing time and subscription behavior data for fund products, calculate the total daily browsing time and subscription interval, and archive and sort them by day to obtain the basic data of user financial behavior time series.

[0017] S102: Based on the time-series baseline of user financial behavior, call the dynamic sliding window algorithm based on user behavior decay factor, set the sliding window to three days, slide the window in time order, and adjust the browsing time and subscription interval data of the window by weighting the decreasing factor to obtain the product attention difference value sequence.

[0018] S103: Based on the product attention difference value sequence, analyze the direction and magnitude of attention changes within three days, mark positive and negative trends and statistically analyze the magnitude of changes to obtain the financial behavior volatility index.

[0019] The dynamic sliding window algorithm based on user behavior decay factor calculates the change in attention by setting the size of the sliding window and the sliding step size. The sliding step size is smaller than the size of the sliding window. When the decay factor λ > 0.7, the window size does not exceed 5 calendar days.

[0020] As a further aspect of the present invention, the specific steps of S2 include:

[0021] S201: Obtain the frequency of hot words in the central bank's policy texts, extract keywords and analyze their frequencies through natural language processing technology, statistically analyze the fluctuations in different time periods, and obtain the hot word frequency curve;

[0022] The natural language processing technology uses statistical analysis and machine learning models to identify words and dynamically track their frequencies.

[0023] S202: Based on the hot word frequency curve and the profit rate change data of competitors, a dynamic weighting method is used to match the data and adjust the influence weights, calculate the deviation between the expected change and the real-time change, and obtain the expected deviation value.

[0024] S203: Based on the comparison between the expected deviation value and the real-time bond product subscription volume curve, analyze the difference between the two and calculate the deviation between market expectations and real-time demand to generate market expectation deviation.

[0025] The frequency of hot words in the central bank's policy text is a dimensionless parameter, and the yield of peer competitors is a percentage. Both are standardized using Z-score to unify the dimensions.

[0026] The weights of the dynamic weighting method are updated in real time through rolling window regression and a machine learning model. The machine learning model uses an LSTM neural network, and the input layer includes raw weight data for 20 time steps.

[0027] As a further aspect of the present invention, the expected deviation value is calculated using the following formula:

[0028] ;

[0029] in, This represents the expected deviation value calculated based on the degree of deviation between the frequency of hot words and the change in returns at time t. This represents the standardized frequency value of the x-th hot word at time t. This represents the average frequency of the x-th hot word within the backtracking window corresponding to time t. This represents the standard deviation of the frequency of the x-th hot word within the retrospective window corresponding to time t. Let represent the change in the rate of return of the z-th competitor at time t, where z∈[1,m] indicates that the selected competitor must meet the standard of having an average daily trading volume of more than 100,000 units. t represents the average change in the returns of all competing products in the same industry at time t, n represents the total number of hot words included in the calculation at the current time, and m represents the total number of competing products in the same industry.

[0030] As a further aspect of the present invention, the specific steps of S3 include:

[0031] S301: Based on the positive and negative signs of the market expectation deviation and its fluctuation characteristics, analyze the product fluctuation range, sort the ranges initially, and generate market deviation range data;

[0032] S302: Based on the market deviation interval data, select a near-end strategy optimization method, set interval matching rules and priority weight parameters, perform a two-way comparison of risk and return on the original risk value and return sequence, add a score of deviation interval label, and output priority optimization ranking;

[0033] S303: Based on the priority optimization sorting, after establishing a sorting mapping relationship according to the sorting weight and product identification data, sort them in descending order of weight value to generate a financial product recommendation queue;

[0034] The positive and negative signs of the market expectation deviation are categorized as positive deviation and negative deviation, respectively.

[0035] The near-end strategy optimization method selects the 30% and 60% quantiles of the volatility index data as the boundaries between low, medium, and high volatility. The quantile thresholds are evaluated by introducing the Monte Carlo simulation method into the original backtesting data to test parameters and model the volatility distribution, assess the stability and discriminative power of the candidate quantiles, and determine that 30% and 60% are the quantile thresholds with better strategy interval division effects.

[0036] As a further aspect of the present invention, the specific steps of S4 include:

[0037] S401: Obtain the sorting result of the wealth management product recommendation queue, extract the annualized rate of return, risk level, investment period and minimum purchase amount feature parameters of the wealth management products in the queue, perform unified dimension, feature encoding and vectorization processing, and construct product feature vector group;

[0038] The annualized rate of return is a percentage, the risk level is a dimensionless parameter, the investment period is days, and the minimum investment amount is yuan. All of these are standardized using Z-score to unify the dimensions.

[0039] S402: Based on the product feature vector group and the user profile vector, the cosine similarity algorithm is used to calculate the similarity between the vectors, and the similarity value is corrected by the risk correction factor of user risk tolerance and product risk level to generate an attribute matching coefficient set;

[0040] The risk correction factor corrects for risk level matching deviation, similarity fluctuation adjustment term, and feature parameter difference attenuation term;

[0041] S403: Call the attribute matching coefficient set, set the quantile interval according to the coefficient distribution law, screen financial products across risk levels to form a combination scheme, and output a customized financial product combination;

[0042] The user profile vector is normalized using the min-max normalization method, and the product feature vector is normalized using the Euclidean space projection method. The user profile vector includes a weighted combination of risk preference dimension, liquidity demand dimension, and return expectation dimension.

[0043] The cross-risk level selection method selects products from different risk level categories by setting a minimum selection ratio threshold for each risk level.

[0044] As a further aspect of the present invention, the risk correction factor is calculated using the following formula:

[0045] ;

[0046] in, This represents the risk adjustment factor between user i and product j. This represents user i's risk tolerance score. This represents the risk level of product j. This represents the highest risk level score. Represents the original similarity value. This represents the average of the original similarity values ​​in the current candidate recommendation set. This represents the standardized value of the k-th feature parameter in the user profile vector. The standardized value of the l-th feature parameter corresponding to the k-th feature in the product feature vector is represented by e, which is a mathematical constant.

[0047] As a further aspect of the present invention, the specific steps of S5 include:

[0048] S501: By monitoring the user's operational behavior regarding the customized financial product portfolio, record three metrics: product click frequency, page dwell time, and detail page visit depth, sort them by timestamp, and construct a behavioral time-series dataset;

[0049] S502: Based on the aforementioned behavioral time series dataset, extract three feature parameters: weekly average access frequency, duration fluctuation coefficient, and product type switching frequency. Use the sliding window method to calculate the feature change rate within the window and generate a behavioral feature evolution set.

[0050] S503: Call the behavioral feature evolution set, apply the ARIMA model to calculate the differential stationary sequence of feature parameters, screen effective features through white noise test, calculate the variance contribution of feature sequence, and obtain the user's financial behavior change score.

[0051] As a further aspect of the present invention, the characteristic change rate is calculated using the following formula:

[0052] ;

[0053] in, The rate of change index represents the nth feature within the kth time window. This represents the observation value of the nth feature at time t. represents the arithmetic mean of the nth feature within the kth time window, and N represents the number of time points included in the sliding window. It represents the range of the nth feature within the kth time window.

[0054] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:

[0055] By employing a dynamic sliding window algorithm to track user financial behavior at high frequency, this approach transforms traditional static behavior analysis into fluctuation index calculations within a continuous time window, effectively capturing patterns of user interest migration. Combining policy hot topics with competitor return data constructs a market expectation deviation index, overcoming the limitations of single-user behavior analysis and achieving cross-validation between the macro market environment and micro-user behavior. A reinforcement learning strategy optimization model automatically adjusts recommendation priorities based on real-time market deviations, enabling marketing strategies to adapt to the environment and addressing the lag issue of traditional fixed-weight recommendation models in dynamic markets. A feature vector matching algorithm maps user profiles to product features in a multi-dimensional space, achieving accurate recommendations through similarity calculations, avoiding the mechanical defects of traditional rule-based matching. A closed-loop feedback mechanism of user behavior monitoring and time-series analysis continuously updates behavioral data, forming a dynamic iterative optimization of marketing strategies and constructing a full-cycle intelligent decision-making chain from data collection to strategy adjustment. This solution achieves breakthroughs in three dimensions: data processing, strategy generation mechanism, and system adaptability, significantly improving the timeliness, accuracy, and environmental adaptability of financial product marketing. Attached Figure Description

[0056] Figure 1 This is a schematic diagram of the workflow of the present invention. Detailed Implementation

[0057] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0058] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0059] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0060] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0061] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0062] Please see Figure 1 This invention provides a marketing method for financial service products based on DeepSeek. The processing flow of this method may include the following steps:

[0063] S1: By collecting the user's financial product browsing time sequence and fund subscription interval sequence, the dynamic sliding window algorithm based on the user behavior decay factor is called to calculate the change in product attention over three consecutive natural days, and generate a financial behavior volatility index.

[0064] S2: By obtaining the frequency of hot words in the central bank's policy texts and the change in yields of peer competitors, the dynamic weighting method is used to compare the demand forecast curve of bond products with the real-time subscription volume curve to obtain the deviation of market expectations.

[0065] S3: Based on the positive and negative signs of market expectation deviation and the numerical range of financial behavior volatility index, a strategy optimization model based on reinforcement learning is used to reorder the recommendation priority of bank wealth management products and generate a wealth management product recommendation queue.

[0066] S4: Based on the ranking results of the financial product recommendation queue, the similarity between the user profile vector and the feature vector of the recommended products is calculated using a feature vector matching algorithm, and a customized financial product portfolio is output.

[0067] S5: By monitoring users' operational behavior in their customized financial product portfolios, update the product browsing time sequence, extract user behavior characteristics, and combine time series analysis methods to obtain a score of changes in user financial behavior;

[0068] The financial behavior volatility index includes changes in the number of product categories monitored, fluctuations in browsing frequency, and trends in behavioral activity. Market expectation deviation specifically includes the range of subscription volume exceeding expectations, the range of subscription volume falling short of expectations, and the range of subscription volume meeting expectations. The wealth management product recommendation queue includes priority product identifiers, product ranking weight values, and recommendation timeliness parameters. Customized financial product portfolios specifically refer to the set of products with the best user preference matching, the risk-return balance index among products, and the validity period setting of customized portfolios. The user financial behavior change score includes the characteristics of changes in browsing behavior stages, the transition signals of subscription behavior stages, and behavioral stability assessment indicators.

[0069] Specifically, the steps in S1 are as follows:

[0070] S101: Collect user browsing time and subscription behavior data for fund products, calculate the total daily browsing time and subscription interval, and archive and sort them by day to obtain the basic data of user financial behavior time series.

[0071] Data on user browsing time and purchase behavior is collected. Browsing behavior data is collected by accessing user terminal operation logs and recording the dwell time on each fund product details page using a page dwell timer mechanism. The browsing time of the same user on different fund product pages within a day is accumulated. For example, if user A browses fund products A, B, and C on May 1st for 120 seconds, 45 seconds, and 95 seconds respectively, the total browsing time for that user on that day is 260 seconds. Purchase behavior data is collected based on the transaction logs of the backend order system, recording the timestamp and corresponding fund identifier for each fund purchase. The time interval between two consecutive valid purchase records is calculated. For example, if a user purchases fund A on May 1st and then purchases fund B on May 4th, the user's purchase interval is 3 days. If a user does not make any purchases on a given day, the record is left blank or set to the maximum interval in days, such as 999. All collected data is aggregated by user dimension to form one record per day. The record includes the user ID, total browsing time, and the number of days between the last purchase date and the current date. This structure is used to build the basic data for the time series of user financial behavior. When archiving, the records are arranged in ascending order by date, so that each user's daily records form a continuous time series. This series is used for subsequent calculations and analysis. For example, if user B browsed fund products for a total of 320 seconds, 210 seconds, and 100 seconds on May 1st, May 2nd, and May 3rd, respectively, and the purchase dates were May 1st, no purchase, and May 3rd, then the corresponding purchase intervals are 0 days, 999 days, and 2 days, respectively. After archiving this time series by day, it can be used for behavioral trend identification and pattern extraction.

[0072] S102: Based on the time-series data of user financial behavior, call the dynamic sliding window algorithm based on the user behavior decay factor, set the sliding window to three days, slide the window in time order, and adjust the browsing time and subscription interval data of the window by weighting the decreasing factor to obtain the product attention difference value sequence.

[0073] Based on the time-series data of user investment behavior, a time-series sliding structure is constructed according to date order. The sliding window is set to three days. Subsequences of three days in length are sequentially extracted from the user's time series, starting from the first day. Within each window, an array of browsing time and an array of purchase intervals are extracted. The data within the window is weighted by setting a behavior decay factor, which is distributed to the data weight of each day in the window in an exponential decay manner. The weights are set to satisfy λ1=1, λ2=λ, λ3=λ. 2 For example, when the decay factor is set to λ=0.8, the range of values ​​for the decay factor λ is iteratively optimized using gradient descent based on historical user behavior datasets to minimize the deviation of the attention difference after behavior decay on the training and validation sets, thereby improving the adaptability of the decay parameter to different user behavior patterns. The weights within 3 days are 1, 0.8, and 0.64 respectively. Then, the browsing time and purchase interval for the three days within the window are multiplied by their corresponding weights and summed to obtain the weighted total browsing time and weighted purchase interval values, denoted as T. view With T gap Then, the difference between the weighted browsing time and the weighted subscription interval of the sliding window is calculated to obtain the product attention difference value, denoted as D=T. view -T gap The difference value is used as the output of the sliding window. The window slides forward one day each time, continuing the above process until a product attention difference value sequence is formed. Each element in this sequence reflects the degree of difference between a user's browsing behavior and purchase behavior within a three-day period. If a user's browsing time from May 1st to May 3rd is 260 seconds, 300 seconds, and 240 seconds respectively, and the purchase intervals are 1 day, 999 days, and 2 days respectively, assuming a decay factor of 0.8, then the weighted browsing time is 260×1 + 300×0.8 + 240×0.64 = 260 + 240 + 153.6 = 653.6 seconds, the weighted purchase interval is 1×1 + 999×0.8 + 2×0.64 = 1 + 799.2 + 1.28 = 801.48 days, and the product attention difference value is D = 653.6 - 801.48 = -147.88 indicates that the purchase behavior lags far behind the browsing behavior in user behavior, suggesting poor conversion of attention.

[0074] S103: Based on the product attention difference value sequence, analyze the direction and magnitude of attention changes within three days, mark positive and negative trends and statistically analyze the magnitude of changes to obtain the financial behavior volatility index.

[0075] The dynamic sliding window algorithm based on user behavior decay factor calculates the change in attention by setting the size of the sliding window and the sliding step size. The sliding step size is smaller than the size of the sliding window. When the decay factor λ > 0.7, the window size does not exceed 5 calendar days.

[0076] Based on the product attention difference value sequence, the trend of difference value changes within a three-day sliding window is analyzed segment by segment in chronological order. A trend judgment benchmark of 0 is set. If there is a sign change between two consecutive difference values, from negative to positive or from positive to negative, it is marked as a trend inflection point. If the difference value continuously increases or decreases and the difference between two adjacent values ​​exceeds a set threshold, it is judged as a trend change. This threshold is set to 100. If the difference value in a certain window is -50 and the difference value in the next window is 60, the difference amplitude is 110, satisfying the condition of a significant trend change. This is marked as a fluctuation event, and the direction of the difference value change is recorded as positive, with a change amplitude of 110. After sequentially scanning the difference value sequence, all trend inflection points and significant fluctuation events are counted to generate a trend change sequence. By calculating the trend switching frequency and the number of fluctuation events per unit time in this sequence, user rationality is obtained. The volatility index for financial behavior is defined as follows: if a user experiences 3 trend reversals and 2 significant volatility events within a seven-day period, the volatility index can be expressed as a trend reversal frequency of 3 / 7 and a significant volatility ratio of 2 / 7. This index is used to assess the instability of changes in a user's financial behavior. For example, if there is a sequence of difference values ​​[100, 120, -50, -80, 150, 170, -30], then the trend reversals occur at positions 2→3 (positive to negative), 4→5 (negative to positive), and 6→7 (positive to negative), totaling 3 reversals. Significant volatility occurs in segments where the difference exceeds 100, namely segments 3→4 (difference from -50 to -80, difference of 30 is negligible), segments 4→5 (-80 to 150, difference of 230), and segments 5→6 (150 to 170, difference of 20 is negligible). Therefore, one significant volatility is recorded.

[0077] Table 1 Examples of User Browsing and Subscription Behaviors

[0078]

[0079] As shown in Table 1, by recording the total daily browsing time and subscription interval of users, it can be used for subsequent calculation of sliding window difference value and fluctuation trend analysis.

[0080] Specifically, the steps of S2 are as follows:

[0081] S201: Obtain the frequency of hot words in the central bank's policy texts, extract keywords and analyze their frequencies through natural language processing technology, and statistically analyze the fluctuations in different time periods to obtain the hot word frequency curve;

[0082] Natural language processing technology uses statistical analysis and machine learning models to identify valuable words and dynamically track their frequencies;

[0083] The process of obtaining the frequency of hot words in the People's Bank of China's policy texts uses the People's Bank of China's quarterly monetary policy implementation report as the input data source. First, each report is structurally decomposed into parts such as title, main body paragraphs, and appendices. The main body is used as the semantic backbone. Natural language processing technology is used to perform lexical analysis on the main body, and the Jieba word segmentation database is called to perform full vocabulary segmentation to form a word unit list. Then, stop words are removed, and valid words with parts of speech of nouns, verbs, and adjectives are selected and retained as candidate sets. Finally, all candidate words are classified by quarter, forming a time-word binary matrix, with each row representing a quarter. The matrix uses a sample size, with each column representing a term dimension. Each cell in the matrix counts the frequency of that term in documents within the current quarter. Frequency standardization is then applied by calculating the mean μ and standard deviation σ of each term's frequency across all quarters. The mean is subtracted from the current quarter's frequency, and the result is divided by the standard deviation to generate the z-score standardized result. For example, the frequency of the hot term "interest rate" in Q1, Q3, and Q4 of 2021 was 0.010, 0.012, and 0.015 respectively, with a corresponding mean of (0.010 + 0.012 + 0.015) / 3 = 0.0123. The standard deviation σ is calculated as follows:

[0084] If the current frequency is 0.013, then the standardized frequency is (0.013-0.0123) / 0.0025=0.28. After processing all hot words, the standardized frequency matrix of all hot words in each quarter is obtained. Based on this, the frequencies of all hot words are sorted by quarter and plotted to form a hot word frequency curve, which reflects the changing trend of the degree of attention of hot words in policy texts in different quarters, as shown in Table 2 below.

[0085] Table 2. Hot Word Frequency and Subscription Data

[0086]

[0087] As shown in Table 2, the frequency of the hot word "inflation" gradually increased from 0.020 in Q1 2021 to 0.025 in Q3, reflecting the increased attention to inflation in the current policy context. The frequency curve declined slightly to 0.023 in Q4.

[0088] S202: Based on the frequency curve of hot words and the profit margin change data of competitors, a dynamic weighting method is used to match the data and adjust the influence weights, calculate the deviation between the expected change and the real-time change, and obtain the expected deviation value.

[0089] Based on the frequency curve of trending keywords and the profit change data of competitors, the process of calculating the deviation from market expectations involves using the frequency values ​​of trending keywords and the profit rates of competitors as input data. First, standardized frequency values ​​of trending keywords in each quarter are obtained according to quarterly segmentation. Then, competitor profit change data is retrieved and differentiated across quarters through classification operations to generate a sequence of profit change values. Subsequently, the frequency of trending keywords and profit change data are paired within the same period to ensure consistency of the data's time window. After pairing, the deviation from market expectations is calculated using the formula:

[0090] ;

[0091] in, This represents the expected deviation value calculated based on the degree of deviation between the frequency of hot words and the change in returns at time t. This represents the standardized frequency value of the x-th hot word at time t. This represents the average frequency of the x-th hot word within the backtracking window corresponding to time t. This represents the standard deviation of the frequency of the x-th hot word within the retrospective window corresponding to time t. Let represent the change in the rate of return of the z-th competitor at time t, where z∈[1,m] indicates that the selected competitor must meet the standard of having an average daily trading volume of more than 100,000 units. represents the average change in the returns of all competing products at time t, n represents the total number of hot words included in the calculation at time t, and m represents the total number of all competing products.

[0092] First, calculate the rolling mean and standard deviation based on the frequency of trending terms. Use data from the first three quarters for this calculation. For example, if the frequency of the trending term "interest rate" in Q1, Q3, and Q4 of 2021 were 0.010, 0.012, and 0.015 respectively, then its mean is:

[0093] ;

[0094] Standard deviation is If the frequency in Q4 2021 is 0.013, then the deviation calculation term is: Then, the variance was calculated based on the changes in competitor returns. Assuming that in Q4 2021, there were three types of competitor returns of 2.12%, 2.11%, and 2.13% respectively, the mean was calculated as follows: The sum of squared deviations is Then calculate the deviation correction factor as follows: Combining the frequency deviation of hot words, the market expectation deviation is calculated as follows: The above calculation steps are performed in the same way on all hot word frequencies, and the market expectation deviation is obtained by summarizing. The larger the value, the lower the market response to policy signals, and the smaller the value, the more consistent the market is with policy signals.

[0095] S203: Based on the comparison between the expected deviation value and the real-time bond product subscription volume curve, analyze the difference between the two and calculate the deviation between market expectations and real-time demand to generate the market expectation deviation.

[0096] The frequency of hot words in the central bank's policy texts is a dimensionless parameter, and the yield of peer competitors is a percentage. Both are standardized using Z-score to unify the dimensions.

[0097] The weights of the dynamic weighting method are updated in real time through rolling window regression and a machine learning model. The machine learning model uses an LSTM neural network, and the input layer contains raw weight data for 20 time steps.

[0098] Based on the comparison between the expected deviation value and the real-time bond product subscription volume curve, the expected deviation data and real-time subscription data first need to be aligned quarterly. The observation period is set to four consecutive quarters. First, the market expected deviation value for each quarter is paired with the corresponding real-time bond product subscription volume value. Taking Q1 to Q4 of 2021 as an example, the corresponding deviation data are {0.0000, 0.0005, 0.0008, 0.000702}, and the corresponding subscription volume is {7.5, 7.9, 8.3, 8.1} billion yuan. Then, the subscription volume change rate for each quarter is calculated, with the change rate set as the current quarter's subscription value minus the previous quarter's... The subscription value in the first quarter is divided by the subscription value in the previous quarter. For example, the change rate between Q2 and Q1 in 2021 is (7.9-7.5) / 7.5=0.0533, the change rate between Q3 and Q2 in 2021 is (8.3-7.9) / 7.9=0.0506, and the change rate between Q4 and Q3 in 2021 is (8.1-8.3) / 8.3=-0.0241. This yields the trend change of the subscription curve. Then, the difference between the deviation and the subscription change rate is calculated. The comparison method involves normalizing both sequences to the [0, 1] interval to ensure data with different dimensions are compared under a unified dimension. The deviation is normalized using the min-max standardization method. Taking the deviation sequence as an example, the maximum value is 0.0008 and the minimum value is 0.0000, so the normalized result is {0.0, 0.625, 1.0, 0.8775}. Similarly, for the normalized subscription change rate, the change rate in the third quarter is {0.0533, 0.0506, -0.0241}, with a maximum of 0.0533 and a minimum of -0.0241, so the normalized result is {1.0, 0.9721, 0.0}. Then, the difference between the quarterly deviation and the normalized subscription change rate is calculated. For example, for Q2 2021... 2021Q3 is 2021Q4 is Finally, the average of these three differences is calculated to obtain the average deviation, which is the deviation between market expectations and real-time demand. The result is (0.375 + 0.0279 + 0.8775) / 3 = 0.4268. A higher value indicates a larger difference between market predictions and real-time subscription responses. The magnitude of the deviation reflects the degree of lag or overreaction in the market's understanding of policy signals. Further analysis is conducted on a quarterly basis using a rolling window. By setting a three-period window, the rolling deviation average is calculated progressively to monitor the gradual deviation between expectations and real-time demand. For example, in Q2-Q4 of 2021, the deviations were 0.375, 0.0279, and 0.8775 respectively, so the rolling average is (0.375 + 0.0279 + 0.8775) / 3 = 0.4268. If subsequent quarters, such as Q1 of 2022, are included, the calculation will be recalculated and replaced. The earliest value is used to track market adaptation by constructing a rolling update model to determine whether the market response in a certain quarter is in a high deviation range. If the rolling mean exceeds 0.4, it is considered "high deviation"; between 0.2 and 0.4, it is considered "medium deviation"; and less than 0.2, it is considered "low deviation". This range is set by the quantile values ​​of the original deviation distribution. Specifically, the 25%, 50%, and 75% quantiles of the distribution are obtained after sorting the average deviation data of the past 40 quarters. Assuming the corresponding quantiles are 0.15, 0.27, and 0.40, then values ​​above the 75th quantile are considered "high deviation", and vice versa. In this example, the market deviation of 0.8775 in Q4 2021 is greater than 0.40, so it is judged as "high deviation". The final market expectation deviation sequence output by quarter serves as the basis for subsequent strategy adjustment analysis.

[0099] Specifically, the steps of S3 are as follows:

[0100] S301: Based on the positive and negative signs of market expectation deviation and the numerical range of financial behavior volatility index, analyze the product volatility range, perform preliminary sorting by range, and generate market deviation range data;

[0101] Based on the sign of market expectation deviation and the numerical range of the financial behavior volatility index, the deviation results for each quarter are first extracted from the quarterly market deviation series, and their sign attribute is determined. If the deviation value is greater than zero, it is recorded as positive; if it is less than zero, it is recorded as negative. Then, two series are established: a positive deviation series and a negative deviation series. The quarterly return volatility of the corresponding financial product in each series is extracted. For each quarterly volatility index, the daily return data of the product within the quarter is used to calculate the standard deviation of the daily returns within the quarter as the volatility index. For example, if the daily returns of a product in Q1 2021 are {0.15%, 0.13%, 0.14%, 0.16%, 0.12%}, then the volatility index is:

[0102] ;

[0103] Subsequently, all quarterly volatility index values ​​were aggregated to form a complete volatility index set. The values ​​were sorted in ascending order, and the 30% and 60% quantile values ​​were extracted as the dividing criteria. The quantile thresholds were evaluated by introducing Monte Carlo simulation into the original backtesting data to test parameters and model volatility distribution, assess the stability and discriminative power of the candidate quantiles, and determine that 30% and 60% were the quantile thresholds with better strategy interval division effects. Let the total volatility set be {0.0120%, 0.0130%, 0.0141%, 0.0160%, 0.0180%, 0.0190%, 0.0210%, 0.0220%}. After sorting, the third term is 0.0141% (approximately the 30th percentile), and the fifth term is 0.0180% (approximately the 60th percentile). Based on this, the low volatility range is defined as (0, 0.0141%), the medium volatility range as (0.0141%, 0.0180%), and the high volatility range as (0.0180%, +∞). Then, the volatility index for each quarter is assigned to a specific range, and all quarters are classified into one of these ranges based on their volatility values. The sign of the deviation from the market expectation is recorded. Based on the joint distribution of positive and negative deviations and volatility ranges, financial products are initially sorted. The sorting process follows the positive... The number of samples in the high volatility range within the offset sample is counted as a primary indicator, medium volatility as a secondary indicator, and low volatility as a tertiary indicator. The negative offset sample follows the opposite sorting logic, with low volatility as a primary indicator. For example, in a product's quarterly performance, there are 3 quarters with positive offsets, 2 in the high volatility range and 1 in the medium volatility range; its ranking priority is (2, 1, 0). Conversely, if another product has 4 quarters with negative offsets, 2 with low volatility, 1 with medium volatility, and 1 with high volatility, its ranking priority is (2, 1, 1). To maintain sorting consistency, the positive and negative offset sorts are standardized to the same structure and then merged. Weights are assigned to the multiple terms in the sorted tuples, such as 0.5 for high volatility, 0.3 for medium volatility, and 0.2 for low volatility. The product's deviation range score is calculated, and the actual score is:

[0104] The score is calculated as follows: (Number of high volatility items × 0.5) + (Number of medium volatility items × 0.3) + (Number of low volatility items × 0.2).

[0105] For example, product A is (2, 1, 0), with a score of 2×0.5+1×0.3+0×0.2=1.3, and product B is (2, 1, 1), with a score of 2×0.2+1×0.3+1×0.5=1.2. After comparison, product A has a higher priority than product B. Finally, based on this score result, all products are initially ranked to form a market deviation interval ranking table, as shown in Table 3.

[0106] Table 3. Ranking of Market Deviation Intervals

[0107]

[0108] Table 3 shows the performance of the four types of products in the deviation range. The score is composed of the number of fluctuations under different offset signs in the combination according to weight, which can be used as the initial basis for subsequent strategy selection.

[0109] S302: Based on market deviation interval data, select a near-end strategy optimization method, set interval matching rules and priority weight parameters, perform a two-way risk-return comparison between the original risk value and return sequence, overlay a scoring function for deviation interval labels, and output a priority optimization ranking.

[0110] Based on the market deviation interval data, the scores of each financial product in Table 3 are first sorted in descending order as the initial interval priority sequence. Then, for the top-ranked products, their original returns and corresponding risk values ​​for the past four quarters are extracted. The risk value is defined as the standard deviation of the return within that quarter, and the return is defined as the total return from the beginning to the end of the quarter divided by the net asset value at the beginning of the quarter. Let the return sequence of Product A for the past four quarters be {2.4%, 2.1%, 2.8%, 2.5%}, and the corresponding risk sequence be {0.95%, 0.88%, 1.03%, 0.92%}. Its average return is calculated to be 2. 0.4 + 2.1 + 2.8 + 2.5 / 4 = 2.45%, the average risk is 0.95 + 0.88 + 1.03 + 0.92 / 4 = 0.945%. This average is used as the basic return-risk performance parameter for the current product. Then, based on the deviation interval labels of the products in the ranking, weighting terms are set for the influence of label offset direction and interval strength on the scoring function. Let positive offset high volatility be weighted at +0.15, medium volatility at +0.10, and low volatility at +0.05, and negative offset at -0.15, -0.10, and -0.05 respectively. Product A's label combination is positive offset high volatility twice and medium volatility once. The weighted total score is 2 × 0.15 + 1 × 0.10 = 0.40. This value is used as the tag scoring factor, combined with the basic risk-return ratio for a composite score. The risk-return ratio is approximately 2.45 / 0.945 ≈ 2.5931, and the final weighted score is 2.5931 + 0.40 = 2.9931. The same process is performed on all products, calculating their scores one by one. Then, they are sorted from highest to lowest score. After completing the two-way risk-return comparison and the optimization steps of the tag scoring function, the resulting ranking is: Product A > Product C > Product D > Product B. If some products have the same score, the ranking is further determined based on the original quarterly rate of return. The standard deviation of volatility serves as a supplementary factor to break the stalemate. It is calculated by comparing the standard deviation of the maximum and minimum quarterly returns. The smaller the difference, the higher the stability and the higher the priority. For example, if the fourth quarter return volatility of product C is {0.90%, 0.89%, 0.91%, 0.88%}, the difference is 0.91-0.88=0.03%, while that of product D is {1.05%, 1.01%, 1.02%, 1.00%}, the difference is 1.05-1.00=0.05%. Therefore, product C has a higher priority than product D, resulting in an optimized product priority ranking, which is used in the next step to build the recommendation queue.

[0111] S303: Priority-based optimization sorting: After establishing a sorting mapping relationship based on sorting weights and product identification data, sort the products in descending order of weight values ​​to generate a financial product recommendation queue.

[0112] The sign of the market expectation deviation is categorized as positive (positive) and negative (negative).

[0113] The near-end strategy optimization method selects the 30th and 60th percentiles of volatility index data as the boundaries between low, medium and high volatility.

[0114] Based on priority-based optimization sorting, and according to the sorting weights and product identification data, the unique identification information of all financial products is first extracted and designated as product numbers {A01, B02, C03, D04}. The optimized sorting scores obtained in the previous step are arranged in descending order and assigned weight values ​​accordingly. The product with the highest score has a weight of 1.00, and the weight values ​​of other products are calculated as a ratio of the score to the maximum value. For example, if product A scores 2.9931 and product C scores 2.8922, then the weight of product C is 2.8922 / 2.9931≈0.9663. This process is repeated to obtain the weight values ​​for all products: A01=1.00, C03=0.9663, D04=0.9450, B02=0.9312. The above product numbers are then... A one-to-one mapping is established between the product number and the weight value, forming the following mapping structure: {A01:1.00, C03:0.9663, D04:0.9450, B02:0.9312}. This structure is then sorted by weight value from largest to smallest, generating the final recommendation queue [A01, C03, D04, B02]. This recommendation queue serves as the final output, directly interacting with the user interface or being used in subsequent strategy model calls. The sorting does not dynamically update until the next evaluation cycle is completed and re-evaluated, at which point it is replaced again. The recommendation queue generated in this step can be directly associated with product numbers to specific issuance information, sales platforms, investment recommendations, and other modules. In subsequent processes, this order is used as the priority order for recommendations or displays.

[0115] Specifically, the steps of S4 are as follows:

[0116] S401: Obtain the sorting results of the wealth management product recommendation queue, extract the annualized rate of return, risk level, investment period and minimum purchase amount of the wealth management products in the queue, and construct a product feature vector group;

[0117] Based on the ranking results of the wealth management product recommendation queue, the system first reads the ranking list output by the system. The list records the unique number and ranking position number of each product. Products are located sequentially from low to high ranking number, and the corresponding product data table is indexed by the product number. Four key parameters are extracted for each product: annualized return, risk level, investment period, and minimum investment amount. Standardized preprocessing operations are performed on each parameter. For example, the annualized return is converted to a uniform unit (e.g., converting 3.5%, 4.0%, and 5.1% to decimals 0.035, 0.040, and 0.051). If the risk level parameter is a level code such as "R1" to "R5", it is mapped to integers 1 to 5 for subsequent vectorization processing. The investment period unit must be uniformly set to "months", such as 90 days. The terms need to be converted to a 3-month term. The minimum purchase amount must be in a unified currency unit and converted to a standard currency unit (e.g., RMB) according to the current exchange rate. After parameter conversion, each product parameter is combined to form a preliminary four-dimensional vector group. Then, a missing value detection and repair strategy is executed for each item in the preliminary vector group. For example, if a product lacks an investment period, the missing value is estimated by comparing the average of products with similar risk levels and return levels, or an interpolation method is used to construct the filler value. Furthermore, the resulting vector group is uniformly normalized according to the feature dimensions. The annualized rate of return is scaled to [0, 1] using an interval scaling method, and the risk level is inversely normalized (e.g., R5 is normalized to 0, R1 is normalized to 1 to match low risk preferences). The investment period and minimum purchase amount are processed using maximum-minimum standardization. The normalized data are then represented as vectors. ,in The normalized rate of return, To normalize the risk level, To normalize the investment cycle, To normalize the minimum investment amount, for example, if the original data for Product A is: annualized return of 5.1%, risk level R3, investment period of 90 days, and minimum investment of 50,000 yuan, after standardization, we set the maximum annualized return to 8% and the minimum to 2%, then... Risk level R3 is mapped to 3, the highest level is 5, and the lowest is 1. Normalization is (5-3) / (5-1)=0.5. The investment period is 90 days, which is 3 months. Let the maximum be 36 months and the minimum be 1 month. After normalization... The minimum purchase amount is set to a maximum of 1 million and a minimum of 1,000 yuan. After normalization, it becomes (50,000 - 1,000) / (1,000,000 - 1,000) = 0.049. The final normalized vector for product A is... All products are constructed using this method to create a set of product feature vectors. The product feature vectors are normalized using Euclidean space projection. Principal component analysis is used to reduce the dimensionality of the product feature vectors to three-dimensional space. After completion, the similarity calculation stage begins.

[0118] Table 4 Original Data Before Product Standardization

[0119]

[0120] As shown in Table 4, the original feature parameters of the product AD will form a feature vector after standardization, which will be used to calculate the user profile matching degree.

[0121] S402: Based on the product feature vector group and the user profile vector, the cosine similarity algorithm is used to calculate the similarity between vectors. The similarity value is corrected by the risk correction factor of user risk tolerance and product risk level to generate an attribute matching coefficient set.

[0122] The risk correction factor corrects for risk level matching bias, similarity fluctuation adjustment term, and characteristic parameter difference attenuation term;

[0123] First, extract user preference vectors from the user profiling module. The multidimensional meanings of this vector correspond to: the user's desired annualized rate of return, risk tolerance, investment period preference, and acceptable minimum purchase amount range. This vector needs to be correlated with the product feature vector. Perform similarity calculation, where i represents the product number.

[0124] To measure vectors and The matching degree between them is determined using cosine similarity as the similarity index, and its formula is as follows:

[0125] ;

[0126] in, Represents a user preference vector. Represents the product feature vector.

[0127] For example, if a user profile vector is And the standardized vector of a certain product is Then its matching degree is calculated as follows:

[0128] ;

[0129] However, since the cosine similarity value must be between [0, 1], a value of 1.008 indicates that due to precision issues, the actual result should be less than or equal to 1. In this example, we take the similarity value to be... In actual calculations, the values ​​should be normalized or corrected.

[0130] Introducing the risk correction factor formula This is used to adjust the basic similarity based on the difference between the product and the user's risk preferences, and to calculate the risk adjustment factor using the formula:

[0131] ;

[0132] in, This represents the risk adjustment factor between user i and product j. This represents user i's risk tolerance score. This represents the risk level of product j. This represents the highest risk level score. Represents the original similarity value. This represents the average of the original similarity values ​​in the current candidate recommendation set. This represents the standardized value of the k-th feature parameter in the user profile vector. The standardized value of the l-th feature parameter corresponding to the k-th feature in the product feature vector is represented by e, which is a mathematical constant.

[0133] The user's risk tolerance is 0.8, the standardized risk level of product A is 0.5, and the maximum risk level is 1 (the maximum value after setting it to level 5).

[0134] ;

[0135] After adjusting the base similarity using a risk correction factor, the final matching score is 1×(1-0.416)≈0.584. This means that the match between product A and the user is 0.584, which is lower than the original base similarity score considering the impact of risk factors.

[0136] Repeat the above calculation process to perform similarity calculations and risk adjustments for all products to be recommended. Assuming we have multiple products, such as products B, C, and D, calculate their matching scores with the user profile using the same steps.

[0137] Based on the final score All products are sorted from highest to lowest score to generate a final recommendation list, in which products with higher scores are given priority for recommendation to users.

[0138] S403: Call the attribute matching coefficient set, set the quantile interval according to the coefficient distribution law, filter financial products across risk levels to form a portfolio plan, and output a customized financial product portfolio;

[0139] User profile vector normalization uses the min-max normalization method, and product feature vector normalization uses the Euclidean space projection method. The user profile vector includes a weighted combination of risk preference dimension, liquidity demand dimension, and return expectation dimension.

[0140] Cross-risk level selection involves setting a minimum selection ratio threshold for each risk level to select products from different risk level categories.

[0141] After calculating the matching score for each product and user profile, the product score set is first... Perform a descending sort to arrange all products from highest to lowest and generate an initial recommendation queue. Next, analyze the distribution characteristics of this score set, calculating the maximum, minimum, mean, and standard deviation. For example, the maximum score is 0.916, the minimum is 0.541, the mean is 0.743, and the standard deviation is 0.112. Then, classify the matching degree levels based on quantiles and calculate the 25th percentile. 50% quantile 75% quantile ,Will As a recommended filtering threshold The process begins by determining if each product's score is greater than or equal to 0.666. If it is less, the product is discarded. For example, product X, with a score of 0.624 (less than 0.666), is discarded. Products with scores above the threshold enter the valid candidate pool. For products in this pool, a risk level stratification process is performed. The risk level parameter for each product is obtained, and its risk segment is determined based on a set range. For example, low risk is defined as a risk level less than 0.3, medium risk as 0.3 to 0.6, and high risk as above 0.6. Candidate products are categorized by risk segment, and the product with the highest matching score in each risk segment is selected as the representative of that segment. For example, product L in the low-risk segment has a score of 0.832 and a risk level of 0.28. It belongs to the low-risk segment and has the highest score in that segment, so it is selected as the low-risk representative. This process is repeated until all risk segments are selected for initial screening. If a segment lacks products that meet the criteria (e.g., no qualified products are found in the medium-risk segment), the search continues to adjacent segments to find products with slightly lower scores but still higher than the threshold. To supplement the product portfolio, for example, products with a risk level of 0.28 but a score of 0.784 are selected to supplement the medium-risk zone. After completing the risk dimension coverage, the core parameters of all selected products are extracted for configuration display, including annualized return, minimum investment amount, and investment period. These parameters are categorized according to preset ranges: annualized return of 3% to 4.5% is categorized as mid-range return; minimum investment amount of less than 1000 yuan is categorized as very low threshold; and investment period of less than 60 days is categorized as short-term. Taking product L as an example, its return is 3.9%, minimum investment amount is 1000 yuan, and investment period is 45 days, so its corresponding tags are "mid-range return" and "very low threshold". The tags "extremely low threshold" and "short cycle" will be used in the visual output module and sorting control module of the subsequent recommendation results. For example, when multiple products have very similar scores (difference less than 0.01), they will be prioritized in ascending order of minimum purchase amount. If they are still equal, the investment cycle and rating will be compared. If there is still no difference, the platform's product popularity parameter will be introduced for the final decision sorting. The final output product set will be determined, and the output quantity will be dynamically adjusted according to the user profile. For example, highly active users with assets exceeding 100,000 yuan can be recommended 10 products, while low-frequency users are limited to 3 to 5 products. Based on this, the system will form a set of product recommendation combinations that meet both high matching degree and cover multiple risk dimensions.

[0142] Specifically, the steps in S5 are as follows:

[0143] S501: By monitoring users' operational behavior with customized financial product portfolios, record three metrics: product click frequency, page dwell time, and detail page visit depth, sort them by timestamp, and construct a behavioral time-series dataset;

[0144] By monitoring users' actions on customized financial product portfolios, when the system receives access records for a user's visit to a specific financial product portfolio page, it first records the click behavior data for each product unit on the page. This step requires specifying the time point corresponding to each click and using that click time as the timestamp of the behavior event. For example, if user ID 1 clicked on a certain product portfolio a total of 35 times during the week of May 1st to May 7th, 2025, the system needs to store 35 records containing the product ID and timestamp. Subsequently, the system obtains the user's cumulative dwell time on the product portfolio page and records it in the behavior database. A session time interval threshold Δt needs to be set; in this embodiment, Δt is set to 600 seconds. If the user's continuous operation time on the page exceeds Δt, it is counted as multiple dwell times in segments. For example, if user ID 1 spends a total of 2100 seconds on the site during the week, and there are 3 instances of disconnection exceeding the threshold, the system should record these as 4 separate behavior records. The system further identifies whether the user accesses the details page of a financial product within the product portfolio, recording the number of times the user enters the details page and the page depth for each visit. Page depth is measured by page path levels. For instance, if a user accesses a product details page from the overview page to the basic information page, and then to the revenue analysis page, the depth is recorded as 3. If the user accesses the details page 10 times during the week, with an average visit depth of 3.2 pages per visit, the behavior records must retain information about each access path and its level. Finally, the click frequency, page dwell time, and details page visit depth are sorted in ascending order according to timestamps to generate a three-dimensional behavioral time-series dataset.

[0145] S502: Based on the behavioral time series dataset, three feature parameters are extracted: weekly average access frequency, duration fluctuation coefficient, and product type switching frequency. The sliding window method is used to calculate the feature change rate within the window to generate a behavioral feature evolution set.

[0146] Based on the behavioral time-series dataset, the system needs to extract three dimensions of features—visit frequency, page dwell time fluctuation, and product switching frequency—weekly using a sliding window approach. The window size is set to 4 weeks, with a step size of 1 week. Each slide performs a feature extraction operation on the data within the window. First, the weekly average visit frequency is calculated. Given weekly click counts of 28, 35, 33, and 37, the weekly average visit frequency is (28+35+33+37) / 4 = 33.25 visits / week. Second, the fluctuation of dwell time is calculated. Given weekly dwell times of 1800, 2100, 1950, and 2250 seconds, the average value is first calculated. seconds, then calculate the standard deviation as The fluctuation coefficient is approximately 161.24 / 2025≈0.08, and the fluctuation is considered to be within the stable range [0, 0.15] based on the interval. Third, when calculating the switching frequency, it is determined whether the clicked products cross categories within the four weeks. If the categories are A, B, and C, and each click is in a different category, it is counted as one switch. For example, if a user records 5, 4, 6, and 3 switches within four weeks, the switching frequency is (5+4+6+3) / 4 = 4.5 times / week. After extracting the basic feature values, the system needs to further measure the rate of change of each feature. This involves normalizing the change of any feature value between two adjacent weeks within a sliding window. For example, if the click frequency is 35 times in week 2 and 33 times in week 3, the rate of change is (33-35) / 35 = -0.057. The final result indicates that the feature's downward trend at this node is 5.7%. The feature change rate is calculated using the formula:

[0147] ;

[0148] in, The rate of change index represents the nth feature within the kth time window. This represents the observation value of the nth feature at time t. represents the arithmetic mean of the nth feature within the kth time window, and N represents the number of time points included in the sliding window. It represents the range of the nth feature within the kth time window.

[0149] Taking click frequency as an example, assuming N=4, the number of clicks is 28, 35, 33, and 37, the average is 33.25, and the standard deviation is 3.86. Then the first term is 3.86 / 33.25≈0.116, the maximum and minimum difference is 37-28=9, and the second term is (9 / 33.25)×(4 / 5)≈0.217. Finally... The rate of change is relatively high, and the entire process outputs a set of behavioral feature evolution.

[0150] S503: Call the behavioral feature evolution set, apply the ARIMA model to calculate the differential stationary sequence of feature parameters, screen effective features through white noise test, calculate the variance contribution of the feature sequence, and obtain the user's financial behavior change score.

[0151] After obtaining the evolution sequence of features, stationarity processing is performed on each feature. The system first performs a first-order differencing operation on each rate of change sequence, where the difference between the rate of change at each time point and the previous time point is used as a new sequence element. For example, if the click frequency change sequence is [+0.25, -0.057, +0.12], the first-order differencing sequence is [-0.307, +0.177]. Next, a white noise test is performed on the differencing sequence to determine if the autocorrelation function of the sequence is entirely within the confidence interval. If it is entirely within the 95% confidence interval... If a feature passes the test, its variance is further extracted from the features that pass the test. The series variance is defined as the contribution of the indicator. If the difference series variances of the three features click frequency, page dwell time, and switching frequency are 0.025, 0.012, and 0.031, respectively, and the total variance is 0.068, then their contributions are 0.025 / 0.068=36.76%, 0.012 / 0.068=17.65%, and 0.031 / 0.068=45.59%, respectively. This contribution is used as the weight.

[0152] To further predict behavioral trends, the system introduces an ARIMA modeling mechanism for the rate-of-change sequences that have passed the stationarity test. Specifically, the system uses the AIC (Akaike Information Criterion) to identify model parameters based on the most recent K periods' feature rate-of-change difference sequences, selecting the optimal order (p, d, q) of the ARIMA model. Each time the sliding window advances one cycle (1 week) and updates the behavioral feature evolution data, the system performs a rolling model update based on the latest sequence and predicts the feature's trend for the next period. For example, if a feature's model parameters are determined to be (p=1, d=1, q=1) after minimizing the AIC, and with a training window length of 8 periods, the system automatically extracts the most recent 8 observation points each week for model fitting and predicts the feature rate of change for the 9th period.

[0153] The system uses RMSE (Root Mean Square Error) as the metric for evaluating prediction accuracy. If the prediction error exceeds a set threshold for two consecutive periods, the system triggers a model reconstruction mechanism to re-identify ARIMA parameters and adjust the window length to ensure the stability of prediction performance. This prediction result can be used for early warning of future behavioral fluctuations, and can also be combined with scoring results to determine whether the evolution trend of abnormal behavior is sustainable.

[0154] The final score is calculated by combining the standardized values ​​of the features, using the following formula:

[0155] ;

[0156] Where S represents the behavioral evaluation value, and n is the number of features. Let i be the standardized value of the i-th feature. The variance contribution of the i-th feature is given by, for example, three standardized values ​​of 0.68, 0.52, and 0.83, respectively. The score would then be:

[0157] S=0.68×0.3676+0.52×0.1765+0.83×0.4559=0.719;

[0158] The results indicate a moderate trend of change in users' recent behavior, and the scoring results are used to classify the status of subsequent financial behavior.

[0159] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A marketing method for financial service products based on DeepSeek, characterized in that: Includes the following steps: S1: By collecting the user's financial product browsing time sequence and fund subscription interval sequence, the dynamic sliding window algorithm based on the user behavior decay factor is called to calculate the change in product attention over three consecutive natural days, and generate a financial behavior volatility index. S2: By obtaining the frequency of hot words in the central bank's policy texts and the change in yields of peer competitors, the dynamic weighting method is used to compare the demand forecast curve of bond products with the real-time subscription volume curve to obtain the deviation of market expectations. S3: Based on the positive and negative signs of the market expectation deviation and the range of financial behavior volatility index values, the recommendation priority of bank wealth management products is reordered using the near-end strategy optimization method to generate a wealth management product recommendation queue. S4: Based on the sorting results of the financial product recommendation queue, the similarity between the user profile vector and the feature vector of the recommended products is calculated using a feature vector matching algorithm, and a customized financial product combination is output. S5: By monitoring the user's operational behavior regarding the customized financial product portfolio, update the product browsing time sequence, extract user behavior characteristics, and combine time series analysis methods to obtain a score for changes in user financial behavior; The specific steps of S1 include: S101: Collect user browsing time and subscription behavior data for fund products, calculate the total daily browsing time and subscription interval, and archive and sort them by day to obtain the basic data of user financial behavior time series. S102: Based on the time-series baseline of user financial behavior, call the dynamic sliding window algorithm based on user behavior decay factor, set the sliding window to three days, slide the window in time order, and adjust the browsing time and subscription interval data of the window by weighting the decreasing factor to obtain the product attention difference value sequence. S103: Based on the product attention difference value sequence, analyze the direction and magnitude of attention changes within three days, mark positive and negative trends and statistically analyze the magnitude of changes to obtain the financial behavior volatility index. The dynamic sliding window algorithm based on user behavior decay factor calculates the change in attention by setting the size of the sliding window and the sliding step size. The sliding step size is smaller than the size of the sliding window. When the decay factor λ > 0.7, the window size does not exceed 5 calendar days.

2. The financial service product marketing method based on DeepSeek according to claim 1, characterized in that, The dynamic sliding window algorithm establishes a parameter relationship between the user behavior decay factor and the sliding window size, so that the weight of the behavior data at different time points within the window decreases over time, dynamically adjusts the sensitivity of user behavior in the time dimension, and finely controls the calculation accuracy of the change in attention. The time series analysis method uses an ARIMA model with time-varying parameters, and its parameter update period is set to T+1 trading days, where T is the time interval between the current trading day and the most recent model training day. The financial behavior volatility index includes changes in the number of product categories monitored, fluctuations in browsing frequency, and trends in behavioral activity. The market expectation deviation specifically includes the range of subscription volume exceeding expectations, the range of subscription volume falling short of expectations, and the range of subscription volume meeting expectations. The wealth management product recommendation queue includes priority product identifiers, product ranking weight values, and recommendation timeliness parameters. The customized financial product portfolio specifically refers to the set of products with the best user preference matching, the risk-return balance index among products, and the validity period setting of the customized portfolio. The user financial behavior change score includes browsing behavior stage change characteristics, subscription behavior stage transition signals, and behavior stability assessment indicators.

3. The marketing method for financial service products based on DeepSeek according to claim 1, characterized in that, The specific steps of S2 include: S201: Obtain the frequency of hot words in the central bank's policy texts, extract keywords and analyze their frequencies through natural language processing technology, statistically analyze the fluctuations in different time periods, and obtain the hot word frequency curve; The natural language processing technology uses statistical analysis and machine learning models to identify words and dynamically track their frequencies. S202: Based on the hot word frequency curve and the profit rate change data of competitors, a dynamic weighting method is used to match the data and adjust the influence weights, calculate the deviation between the expected change and the real-time change, and obtain the expected deviation value. S203: Based on the comparison between the expected deviation value and the real-time bond product subscription volume curve, analyze the difference between the two and calculate the deviation between market expectations and real-time demand to generate market expectation deviation. The frequency of hot words in the central bank's policy text is a dimensionless parameter, and the yield of peer competitors is a percentage. Both are standardized using Z-score to unify the dimensions. The weights of the dynamic weighting method are updated in real time through rolling window regression and a machine learning model. The machine learning model uses an LSTM neural network, and the input layer includes raw weight data for 20 time steps.

4. The marketing method for financial service products based on DeepSeek according to claim 3, characterized in that, The expected deviation value is calculated using the following formula: ； in, This represents the expected deviation value calculated based on the degree of deviation between the frequency of hot words and the change in returns at time t. This represents the standardized frequency value of the x-th hot word at time t. This represents the average frequency of the x-th hot word within the backtracking window corresponding to time t. This represents the standard deviation of the frequency of the x-th hot word within the retrospective window corresponding to time t. Let represent the change in the rate of return of the z-th competitor at time t, where z∈[1,m] indicates that the selected competitor must meet the standard of having an average daily trading volume of more than 100,000 units. t represents the average change in the returns of competitors at time t, n represents the total number of hot words included in the calculation at the current time, and m represents the total number of competitors.

5. The marketing method for financial service products based on DeepSeek according to claim 3, characterized in that, The specific steps of S3 include: S301: Based on the positive and negative signs of the market expectation deviation and its fluctuation characteristics, analyze the product fluctuation range, sort the ranges initially, and generate market deviation range data; S302: Based on the market deviation interval data, select a near-end strategy optimization method, set interval matching rules and priority weight parameters, perform a two-way comparison of risk and return on the original risk value and return sequence, add a score of deviation interval label, and output priority optimization ranking; S303: Based on the priority optimization sorting, after establishing a sorting mapping relationship according to the sorting weight and product identification data, sort them in descending order of weight value to generate a financial product recommendation queue; The positive and negative signs of the market expectation deviation are categorized as positive deviation and negative deviation, respectively. The near-end strategy optimization method selects the 30% and 60% quantiles of the volatility index data as the boundaries between low, medium, and high volatility. The quantile thresholds are evaluated by introducing the Monte Carlo simulation method into the original backtesting data to test parameters and model the volatility distribution, assess the stability and discriminative power of the candidate quantiles, and determine that 30% and 60% are the quantile thresholds with better strategy interval division effects.

6. The marketing method for financial service products based on DeepSeek according to claim 5, characterized in that, The specific steps of S4 include: S401: Obtain the sorting result of the wealth management product recommendation queue, extract the annualized rate of return, risk level, investment period and minimum purchase amount feature parameters of the wealth management products in the queue, perform unified dimension, feature encoding and vectorization processing, and construct product feature vector group; The annualized rate of return is a percentage, the risk level is a dimensionless parameter, the investment period is days, and the minimum investment amount is yuan. All of these are standardized using Z-score to unify the dimensions. S402: Based on the product feature vector group and the user profile vector, the cosine similarity algorithm is used to calculate the similarity between the vectors, and the similarity value is corrected by the risk correction factor of user risk tolerance and product risk level to generate an attribute matching coefficient set; The risk correction factor corrects for risk level matching deviation, similarity fluctuation adjustment term, and feature parameter difference attenuation term; S403: Call the attribute matching coefficient set, set the quantile interval according to the coefficient distribution law, screen financial products across risk levels to form a combination scheme, and output a customized financial product combination; The user profile vector is normalized using the min-max normalization method, and the product feature vector is normalized using the Euclidean space projection method. The user profile vector includes a weighted combination of risk preference dimension, liquidity demand dimension, and return expectation dimension. The cross-risk level selection method selects products from differentiated risk level categories by setting a minimum selection ratio threshold for each risk level.

7. The marketing method for financial service products based on DeepSeek according to claim 6, characterized in that, The risk correction factor is calculated using the following formula: ； in, This represents the risk adjustment factor between user i and product j. This represents user i's risk tolerance score. This represents the risk level of product j. This represents the highest risk level score. Represents the original similarity value. This represents the average of the original similarity values ​​in the current candidate recommendation set. This represents the standardized value of the k-th feature parameter in the user profile vector. The standardized value of the l-th feature parameter corresponding to the k-th feature in the product feature vector is represented by e, which is a mathematical constant.

8. The marketing method for financial service products based on DeepSeek according to claim 6, characterized in that, The specific steps of S5 include: S501: By monitoring the user's operational behavior regarding the customized financial product portfolio, record three metrics: product click frequency, page dwell time, and detail page visit depth, sort them by timestamp, and construct a behavioral time-series dataset; S502: Based on the aforementioned behavioral time series dataset, extract three feature parameters: weekly average access frequency, duration fluctuation coefficient, and product type switching frequency. Use the sliding window method to calculate the feature change rate within the window and generate a behavioral feature evolution set. S503: Call the behavioral feature evolution set, apply the ARIMA model to calculate the differential stationary sequence of feature parameters, screen effective features through white noise test, calculate the variance contribution of feature sequence, and obtain the user's financial behavior change score.

9. The marketing method for financial service products based on DeepSeek according to claim 8, characterized in that, The rate of change of the aforementioned characteristic is calculated using the following formula: ； in, The rate of change index represents the nth feature within the kth time window. This represents the observation value of the nth feature at time t. represents the arithmetic mean of the nth feature within the kth time window, and N represents the number of time points included in the sliding window. It represents the range of the nth feature within the kth time window.

Images