A financial business risk control management system based on big data model

Abstract

The present invention belongs to the field of big data technology, and specifically refers to a financial business risk control management system based on a big data model, including: a technical information collection module, an emotional information collection module, a fund misappropriation risk assessment module, a user interaction module, a financial analysis module, a behavior analysis module and a comprehensive analysis module. The fund misappropriation risk assessment module uses big data public opinion analysis and investment habit analysis to effectively assess the probability that a customer will misappropriate a loan for risky investment, and provides a basis for targeted post-loan fund supervision to prevent fund misappropriation. The behavior analysis module uses a fusion integrated neural network model to prevent AI fraud by analyzing biological behavior characteristics such as the typing rhythm and mouse trajectory of the customer during the application process. The present invention organically combines big data technology, artificial intelligence technology and statistical analysis to achieve multimodal, explainable and intelligent financial business risk control.

Description

Technical Field

[0001] The present invention belongs to the field of big data technology, and specifically refers to a financial business risk control management system based on a big data model. Background Art

[0002] The financial business risk control management system is a system used to monitor, evaluate and manage financial business risks. It determines the risk level of financial business by analyzing and processing large amounts of data. In the risk control of financial business, credit business is the most typical. For this reason, the industry has developed a variety of risk control systems to capture potential risks in credit, including systems based on credit scoring systems and systems based on artificial intelligence technology. However, the problem with these systems is that systems based on credit scoring systems are difficult to solve the increasingly rampant AI fraud problem in recent years and cannot distinguish between real customers and disguised fake customers. Systems based on artificial intelligence technology can solve the AI ​​fraud problem, but existing systems based on artificial intelligence technology are mostly based on user portrait data. As fraudsters reverse engineer the system, the identification performance of existing systems based on artificial intelligence technology has also declined. At present, there is an urgent need for AI fraud identification methods that are difficult to be reverse engineered. Due to its inherent black box nature, it is impossible to make a reasonable explanation for its decision. It can only play an auxiliary role in fields such as finance where the rationality of decision-making is relatively high, and it is impossible to explain the reasons for the decision. At present, investors are increasingly speculative and institutions’ response to loan misappropriation is mainly post-loan supervision and ex post remedy. The existing financial business risk control management system is also difficult to identify the risk of loan misappropriation before the loan. Summary of the invention

[0003] The present invention overcomes the shortcomings of the prior art and provides a financial business risk control management system based on a big data model. The existing fund misappropriation risk assessment is based on the customer's personal attributes, and fails to take into account the impact of changes in overall investment sentiment on personal thoughts. In addition, the existing investment sentiment index is based on the analysis of the text of each investment forum, which is obviously no longer compatible with the current public information acquisition habits. For this reason, the fund misappropriation risk assessment module uses big data public opinion analysis of the streaming media business platform and investment habit analysis of customers to effectively assess the probability that customers will misappropriate loans for risky investments, providing a basis for targeted post-loan fund supervision to prevent fund misappropriation. In order to prevent reverse engineering, the behavior analysis module uses a fusion integrated neural network model to analyze The bio-behavioral characteristics such as the typing rhythm and mouse trajectory of the customer during the application process are analyzed to prevent AI fraud. Since the bio-behavioral characteristics are dynamic and real-time, they are difficult to crack or imitate. The fusion integrated neural network model combines the ideas of integrated learning and model fusion. The three branches of the model are heterogeneous classifiers while sharing features, and the accuracy is improved by error offset. The financial analysis module combines Benford distribution analysis and association rule mining in mathematical models. Association rule mining analyzes the customer's credit and financial affordability separately, which effectively improves the decision-making rationality and interpretability of the system. The present invention organically combines big data technology, artificial intelligence technology and statistical analysis to achieve multimodal, interpretable and intelligent financial business risk control.

[0004] The financial business risk control management system based on the big data model includes: a technical information collection module, an emotional information collection module, a fund misappropriation risk assessment module, a user interaction module, a financial analysis module, a behavior analysis module and a comprehensive analysis module;

[0005] The technical information collection module collects investment index history records and inputs them into the fund misappropriation risk assessment module, wherein the investment index history records include the Shanghai and Shenzhen 300 Index history records, the commodity housing sales price index history records and the futures comprehensive index history records; the emotional information collection module collects finance-related video information of the streaming media service platform to calculate the stock market investment sentiment index, the real estate investment sentiment index and the futures investment sentiment index, and inputs them into the fund misappropriation risk assessment module;

[0006] The user interaction module prompts the customer to upload the information required for the credit application when applying for a loan. The information uploaded by the customer includes the loan term, loan amount, bank flow, proof of address, personal credit report and other personal information, and records the customer's mouse trajectory and typing rhythm data at the same time. The user interaction module inputs the bank flow into the fund misappropriation risk assessment module, inputs the information uploaded by the customer into the financial analysis module, and inputs the customer's mouse trajectory and typing rhythm data into the behavior analysis module;

[0007] The behavior analysis module constructs a fusion integrated neural network model and uses the fusion integrated neural network model to obtain the probability that the customer is a robot based on the customer's mouse trajectory and typing rhythm data and transmits it to the comprehensive analysis module;

[0008] The financial analysis module obtains the material fraud probability, credit default probability and relative default probability based on the information uploaded by the customer and transmits them to the comprehensive analysis module;

[0009] The fund misappropriation risk assessment module obtains the loan misappropriation index to the comprehensive analysis module based on the information provided by the technical information collection module and the emotional information collection module and the bank flow;

[0010] The comprehensive analysis module determines whether the application is approved and how to respond.

[0011] Furthermore, the process of the emotional information collection module collecting the finance-related video information of the streaming media service platform to calculate the stock market investment sentiment index, the real estate investment sentiment index and the futures investment sentiment index specifically includes the following steps:

[0012] Step S1: Select videos containing finance tags from all videos on the streaming media service platform, and obtain the playback volume of the videos containing finance tags;

[0013] Step S2: sort the videos containing the finance tag in the last month from high to low according to the number of views, and construct a collection of finance videos;

[0014] Step S3: Download the videos in the financial video collection and obtain the release time, number of bullet screens, number of comments, number of favorites, number of shares, number of followers and number of likes of each video in the financial video collection;

[0015] Step S4: TF-IDF matrixing of text and calculating sentiment value of annotated keywords;

[0016] Step S5: extracting acoustic features: locating the time interval in which the keywords appear in the audio and sorting the keywords in the order of their appearance in the audio, extracting the highest pitch of the audio in the time interval as the keyword pitch, and performing maximum-minimum normalization on the keyword pitch of each keyword;

[0017] Step S6: Calculate the video emotion index of each video in the financial video set. The video emotion index formula is as follows:

[0018] ;

[0019] in, represents the video sentiment index, n represents the number of keywords in the video, and i represents the chronological order of the keywords in the video. Represents the keyword pitch, Represents the TF-IDF value of the keyword in the TF-IDF matrix. Represents the sentiment value of the keyword;

[0020] Step S7: Classify the videos in the financial video set into a stock market video set, a real estate video set, and a futures video set according to the labels, while keeping their order;

[0021] Step S8: Calculate the investment sentiment index of the stock market video set as the stock market investment sentiment index. The calculation formula of the investment sentiment index is as follows:

[0022] ;

[0023] ;

[0024] ;

[0025] ;

[0026] in, Represents the sequence number of the video. Representative The spread of the video, Representative Number of shares of videos, Representative The number of video views, Representative Number of followers of the publisher of the video, Representative The discussion intensity of the video, Representative Number of comments in the video, Representative Number of comments on videos, Representative The emotional recognition weight of each video, Represents the total number of videos, Representative Video sentiment index of videos, Representative Number of favorites of videos, Representative Number of favorites of videos, Representative The number of days since the video was published;

[0027] Step S9: Apply the formula in step S8 to process the real estate video set and the futures video set to calculate the real estate investment sentiment index and the futures investment sentiment index.

[0028] Furthermore, the process of the fund misappropriation risk assessment module obtaining the loan misappropriation index according to the information provided by the technical information collection module and the emotional information collection module and the bank flow includes the following steps:

[0029] Step A1: The fund misappropriation risk assessment module extracts the customer's stock investment records, real estate transaction records and futures investment records from the bank flow according to the transaction remarks in the bank flow;

[0030] Step A2: Calculate the customer's stock investment times in the most recent month, real estate transactions times in the most recent month, futures investments times in the most recent month, average stock investment times per month in the most recent year, average real estate transactions times per month in the most recent year, and average futures investments times per month in the most recent year based on the customer's stock investment records, real estate transactions times, and futures investments times in the most recent month;

[0031] Step A3: Calculate the historical volatility of the CSI 300 Index in the most recent month, the monthly average historical volatility of the CSI 300 Index in the most recent six months, the historical volatility of the commercial housing sales price index in the most recent month, the monthly average historical volatility of the commercial housing sales price index in the most recent six months, the historical volatility of the futures composite index in the most recent month, and the monthly average historical volatility of the futures composite index in the most recent six months based on the historical records of the CSI 300 Index, the historical records of the commercial housing sales price index, and the historical records of the futures composite index;

[0032] Step A4: Calculate the loan misappropriation index. The calculation formula of the loan misappropriation index is as follows:

[0033] ;

[0034] ;

[0035] ;

[0036] ;

[0037] in, Represents the amount of stock loan misappropriation, Represents the amount of misappropriation of real estate loans, Represents the amount of futures loan misappropriation, Represents the stock market investment sentiment index, Represents the real estate investment sentiment index, Represents the futures investment sentiment index, Represents the average number of stock investments per month in the most recent year, Represents the average number of real estate transactions per month in the past year. Represents the average number of futures investments per month in the most recent year. Represents the number of stock investments in the most recent month. Represents the number of real estate transactions in the most recent month. Represents the number of futures investments in the most recent month. Represents the volatility of the Shanghai and Shenzhen 300 Index in the most recent month. Represents the average monthly historical volatility of the CSI 300 Index in the last six months. Represents the historical volatility of the commodity housing sales price index in the most recent month, Represents the historical volatility of the monthly average commodity housing sales price index in the last six months, Represents the historical volatility of the futures composite index in the most recent month. Represents the historical volatility of the monthly average futures composite index in the last six months. stands for loan misappropriation index.

[0038] Furthermore, the behavior analysis module constructs a fusion integrated neural network model comprising the following steps:

[0039] Step B1: Constructing behavior recognition dataset;

[0040] Step B2: construct a fusion integrated neural network model;

[0041] Step B3: Use the behavior recognition dataset to train the fusion ensemble neural network model.

[0042] Furthermore, the step B1 specifically includes the following steps:

[0043] Step B11: collecting past mouse trajectory and typing rhythm data;

[0044] Step B12: extracting a mouse coordinate time series, a mouse event time series and a key time series from the past mouse trajectory and typing rhythm data, wherein the mouse coordinate time series records the coordinates of the mouse on the screen along the time axis, the mouse event time series records the mouse click events along the time axis, and the key time series records which keyboard keys are pressed at each moment along the time axis;

[0045] Step B13: superimpose the mouse coordinate time series and the mouse event time series to generate a multi-channel mouse time series, aggregate the multi-channel mouse time series and the key time series to generate a positive sample of the behavior recognition data set, and add a true or false mark to the positive sample of the behavior recognition data set, and the true or false mark value is 1;

[0046] Step B14: randomly generate multi-channel mouse time series and key time series, and summarize them into negative samples of the behavior recognition data set. The true and false label value of the negative samples of the behavior recognition data set is 0;

[0047] Step B15: Mix the negative samples of the behavior recognition data set with the positive samples of the behavior recognition data set to generate a behavior recognition data set.

[0048] Furthermore, the step B2 specifically includes the following steps:

[0049] Step B21: construct a concatenation layer, a fifth fully connected layer, and a sixth fully connected layer and connect them in sequence to construct a fusion feature branch, which performs feature fusion on its input and obtains a first robot probability;

[0050] Step B22: construct a one-dimensional convolutional layer, a Flatten layer, a first fully connected layer, and a second fully connected layer, and connect them in sequence, connect the Flatten layer to the concatenation layer, and construct a mouse trajectory feature branch. The mouse trajectory feature branch extracts the comprehensive mouse features in its input and shares them to the fusion feature branch, and finally obtains the second robot probability;

[0051] Step B23: constructing an LSTM layer, a first fully connected layer, and a second fully connected layer, and connecting them in sequence, connecting the LSTM layer to the concatenation layer, constructing a typing rhythm feature branch, wherein the typing rhythm feature branch extracts the typing rhythm feature in its input and shares it to the fusion feature branch, and finally derives the third robot probability;

[0052] Step B24: Specify the one-dimensional convolution layer and the LSTM layer as the input layer, specify the second fully connected layer, the sixth fully connected layer, and the fourth fully connected layer as the output layer, specify the optimizer as Adam, specify the binary cross entropy as the loss function, and build a fusion integrated neural network model.

[0053] Furthermore, the step B3 specifically includes the following steps:

[0054] Step B31: Divide the behavior recognition dataset into a training set and a test set in a ratio of 8:2;

[0055] Step B32: Specify the multi-channel mouse time series as the input of the one-dimensional convolutional layer, the key time series as the input of the LSTM layer, the true and false labels as the target outputs of the second fully connected layer, the sixth fully connected layer, and the fourth fully connected layer, and train the fusion integrated neural network model.

[0056] Furthermore, the process of the financial analysis module obtaining the probability of material fraud and the probability of credit default based on the information uploaded by the customer specifically includes the following steps:

[0057] Step F1: extract the community name from the address certificate, query the transaction prices of all real estates in the community in the last month based on the community name, and average them to obtain the average housing price in the last month;

[0058] Step F2: Extract the deposit balance and average monthly income in the most recent year from the bank statement;

[0059] Step F3: Extract the first digit of all amounts in bank statements and credit reports, and create a first digit frequency table;

[0060] Step F4: Calculate the KS value between the first digit frequency table and the Benford distribution as the probability of material fraud;

[0061] Step F5: Calculate the customer's expected repayment ability. The calculation formula of the expected repayment ability is as follows:

[0062] ;

[0063] in, Represents expected repayment capacity, Represents the average housing price in the most recent month. Represents the deposit balance, Represents the average monthly income in the most recent year, The number of months representing the loan term;

[0064] Step F6: Collect past credit records and build a credit multidimensional data warehouse;

[0065] Step F7: constructing an Apriori association rule mining model using the credit multidimensional data warehouse;

[0066] Step F8: Use the Apriori association rule mining model to calculate the credit default probability and relative default probability.

[0067] Furthermore, the process of the comprehensive analysis module determining whether the application is approved and how to respond specifically includes the following steps:

[0068] Step C1: Setting the loan misappropriation index threshold, fraud probability threshold, and default probability threshold;

[0069] Step C2: If the sum of the probability of forged materials and the probability that the customer is a robot is greater than the threshold of the probability of forgery, the application will be rejected and a reply will be given: "Hello, as your identity and application materials need further verification, please bring relevant documents to go offline for processing";

[0070] Step C3: If the credit default probability is greater than the default probability threshold, the application will be rejected and a reply will be given: "Hello, after evaluation, your loan application has been rejected due to poor credit record";

[0071] Step C4: If the relative default probability is greater than the default probability threshold, the application will be rejected and the reply will be “Hello, after evaluation, your loan application has been rejected because the loan amount exceeds your financial capacity. We recommend that you reconsider the loan amount.”

[0072] Step C5: If the application is not rejected in steps C2 to C4, the application is approved;

[0073] Step C6: If the loan misappropriation index is greater than the loan misappropriation index threshold, the customer will be marked as a fund misappropriation supervision subject and replied with "Hello, according to regulatory requirements, we will monitor the use of loan funds. Please ensure that the loan funds are not used for high-risk investments to avoid violating relevant regulations."

[0074] The beneficial effects of the present invention are as follows:

[0075] (1) The fund misappropriation risk assessment module of the present invention uses big data public opinion analysis and investment habit analysis to effectively assess the probability of customers misappropriating loans for risky investments, providing a basis for targeted post-loan fund supervision to prevent fund misappropriation. The behavior analysis module uses a fusion integrated neural network model to prevent AI fraud by analyzing the biological behavior characteristics of customers such as typing rhythm and mouse trajectory during the application process. The fusion integrated neural network model combines the ideas of integrated learning with model fusion. The three branches of the model are heterogeneous classifiers while sharing features, and the accuracy is improved by error compensation. The financial analysis module combines Benford distribution analysis and association rule mining in mathematical models. Association rule mining analyzes the customer's credit and financial affordability separately, effectively improving the decision rationality and interpretability of the system. The present invention organically combines big data technology, artificial intelligence technology and statistical analysis to achieve multimodal, interpretable and intelligent financial business risk control.

[0076] (2) In existing risk control management systems, credit and financial affordability are often analyzed together. Considering that the correlation between credit and financial affordability is not as strong as people think, and may even be negatively correlated, the Apriori association rule mining model of the present invention analyzes credit and financial affordability separately, which improves the rationality of decision-making and overcomes the technical bias that credit and financial affordability are strongly correlated.

[0077] (3) The comprehensive analysis module of the present invention allows the system to adjust its sensitivity to various indicators based on self-set thresholds, and provides reasonable rejection reasons for different situations. This is something that systems based on artificial intelligence technology cannot do. Compared with unexplainable black box judgments, the design of the comprehensive analysis module of the present invention is more reasonable in decision-making and can better demonstrate humanistic care.

[0078] (4) The difference between the emotional information collection module of the present invention and the prior art lies in that it introduces multimodal information in the video, breaks through the text form limitation of the traditional investment index, conforms to the changes in the current information browsing habits of investors, and can more sensitively and accurately measure the changes in investment sentiment. The loan misappropriation index of the fund misappropriation risk assessment module conducts a comprehensive analysis of investment habits, the volatility of each segmented investment market and the changes in market investment sentiment, thereby effectively reducing the risk of fund misappropriation.

[0079] (5) In response to the increasingly serious AI fraud problem in recent years, the present invention introduces the design concept of statistical analysis and behavioral event analysis, and uses Benford analysis plus a fusion integrated neural network model to analyze the mathematical statistical characteristics and biological behavioral characteristics in the application process, effectively solving the pain point that existing financial business risk control cannot cope with AI fraud. In addition, due to the dynamic and real-time nature of biological behavioral characteristics, it can effectively counter the reverse engineering of the fusion integrated neural network model. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] Figure 1 This is a module diagram of a financial business risk control management system based on a big data model proposed by the present invention;

[0081] Figure 2 A schematic diagram of the fusion integrated neural network model proposed in the present invention;

[0082] Figure 3 It is the flowchart of the comprehensive analysis module;

[0083] Figure 4 It is a schematic diagram of the key-pressing time sequence;

[0084] Figure 5 Schematic diagram of multi-channel mouse time series.

[0085] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention. DETAILED DESCRIPTION

[0086] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the examples given are only used to explain the present invention, and are not used to limit the scope of the present invention.

[0087] Example 1, see Figure 1 , the present invention provides a financial business risk control management system based on a big data model, including: a technical information collection module, an emotional information collection module, a fund misappropriation risk assessment module, a user interaction module, a financial analysis module, a behavior analysis module and a comprehensive analysis module;

[0088] The technical information collection module collects investment index history records and inputs them into the fund misappropriation risk assessment module, wherein the investment index history records include the Shanghai and Shenzhen 300 Index history records, the commodity housing sales price index history records and the futures comprehensive index history records; the emotional information collection module collects finance-related video information of the streaming media service platform to calculate the stock market investment sentiment index, the real estate investment sentiment index and the futures investment sentiment index, and inputs them into the fund misappropriation risk assessment module;

[0089] The user interaction module prompts the customer to upload the information required for the credit application when applying for a loan. The information uploaded by the customer includes the loan term, loan amount, bank flow, proof of address, personal credit report and other personal information, and records the customer's mouse trajectory and typing rhythm data at the same time. The user interaction module inputs the bank flow into the fund misappropriation risk assessment module, inputs the information uploaded by the customer into the financial analysis module, and inputs the customer's mouse trajectory and typing rhythm data into the behavior analysis module;

[0090] The behavior analysis module constructs a fusion integrated neural network model and uses the fusion integrated neural network model to obtain the probability that the customer is a robot based on the customer's mouse trajectory and typing rhythm data and transmits it to the comprehensive analysis module;

[0091] The financial analysis module obtains the material fraud probability, credit default probability and relative default probability based on the information uploaded by the customer and transmits them to the comprehensive analysis module;

[0092] The fund misappropriation risk assessment module obtains the loan misappropriation index to the comprehensive analysis module based on the information provided by the technical information collection module and the emotional information collection module and the bank flow;

[0093] The comprehensive analysis module determines whether the application is approved and how to respond.

[0094] Embodiment 2: Based on the above embodiment, the process of the emotional information collection module collecting financial related video information of the streaming media service platform to calculate the stock market investment sentiment index, the real estate investment sentiment index and the futures investment sentiment index specifically includes the following steps:

[0095] Step S1: Select a video containing a finance tag from all videos on the streaming media service platform, and obtain playback of the video containing the finance tag;

[0096] Step S2: sort the videos containing the finance tag in the last month from high to low according to the number of views, and construct a collection of finance videos;

[0097] Step S3: Download the videos in the financial video collection and obtain the release time, number of bullet screens, number of comments, number of favorites, number of shares, number of followers and number of likes of each video in the financial video collection. In this embodiment, step S3 is implemented using Feigua data;

[0098] Step S4: TF-IDF matrixing of text and calculating sentiment values ​​of annotated keywords: extract the audio of the videos in the financial video collection, convert it into text and generate analysis text, extract keywords in the analysis text and calculate the TF-IDF matrix of the analysis text, annotate the sentiment values ​​of keywords according to the "Chinese Financial Sentiment Dictionary", where positive sentiment is 1, neutral sentiment is 0, and negative sentiment is -1;

[0099] Step S5: extracting acoustic features: locating the time interval in which the keywords appear in the audio and sorting the keywords in the order of their appearance in the audio, extracting the highest pitch of the audio in the time interval as the keyword pitch, and performing maximum-minimum normalization on the keyword pitch of each keyword;

[0100] Step S6: Calculate the video emotion index of each video in the financial video set. The video emotion index formula is as follows:

[0101] ;

[0102] in, represents the video sentiment index, n represents the number of keywords in the video, and i represents the chronological order of the keywords in the video. Represents the keyword pitch, Represents the TF-IDF value of the keyword in the TF-IDF matrix. The sentiment value representing the keyword, in this embodiment, the MapReduce programming model is used to calculate the sentiment index of each video in the financial video set;

[0103] Step S7: Classify the videos in the financial video set into a stock market video set, a real estate video set, and a futures video set according to the labels, while keeping their order;

[0104] Step S8: Calculate the investment sentiment index of the stock market video set as the stock market investment sentiment index. The calculation formula of the investment sentiment index is as follows:

[0105] ;

[0106] ;

[0107] ;

[0108] ;

[0109] in, Represents the sequence number of the video. Representative The spread of the video, Representative Number of shares of videos, Representative The number of video views, Representative Number of followers of the publisher of the video, Representative The discussion intensity of the video, Representative Number of comments in the video, Representative Number of comments on videos, Representative The emotional recognition weight of each video, Represents the total number of videos, Representative Video sentiment index of videos, Representative Number of favorites of videos, Representative Number of favorites of videos, Representative The number of days since the video was published;

[0110] Step S9: Apply the formula in step S8 to process the real estate video set and the futures video set to calculate the real estate investment sentiment index and the futures investment sentiment index.

[0111] Embodiment 3: This embodiment is based on embodiment 2. The process of the fund misappropriation risk assessment module obtaining the loan misappropriation index according to the information provided by the technical information collection module and the emotional information collection module and the bank flow includes the following steps:

[0112] Step A1: The fund misappropriation risk assessment module extracts the customer's stock investment records, real estate transaction records and futures investment records from the bank flow according to the transaction remarks in the bank flow;

[0113] Step A2: Calculate the customer's stock investment times in the most recent month, real estate transactions times in the most recent month, futures investments times in the most recent month, average stock investment times per month in the most recent year, average real estate transactions times per month in the most recent year, and average futures investments times per month in the most recent year based on the customer's stock investment records, real estate transactions times, and futures investments times in the most recent month;

[0114] Step A3: Calculate the historical volatility of the CSI 300 Index in the most recent month, the monthly average historical volatility of the CSI 300 Index in the most recent six months, the historical volatility of the commercial housing sales price index in the most recent month, the monthly average historical volatility of the commercial housing sales price index in the most recent six months, the historical volatility of the futures composite index in the most recent month, and the monthly average historical volatility of the futures composite index in the most recent six months based on the historical records of the CSI 300 Index, the historical records of the commercial housing sales price index, and the historical records of the futures composite index;

[0115] Step A4: Calculate the loan misappropriation index. The calculation formula of the loan misappropriation index is as follows:

[0116] ;

[0117] ;

[0118] ;

[0119] ;

[0120] in, Represents the amount of stock loan misappropriation, Represents the amount of misappropriation of real estate loans, Represents the amount of futures loan misappropriation, Represents the stock market investment sentiment index, Represents the real estate investment sentiment index, Represents the futures investment sentiment index, Represents the average number of stock investments per month in the most recent year, Represents the average number of real estate transactions per month in the past year. Represents the average number of futures investments per month in the most recent year. Represents the number of stock investments in the most recent month. Represents the number of real estate transactions in the most recent month. Represents the number of futures investments in the most recent month. Represents the volatility of the Shanghai and Shenzhen 300 Index in the most recent month. Represents the average monthly historical volatility of the CSI 300 Index in the last six months. Represents the historical volatility of the commodity housing sales price index in the most recent month, Represents the historical volatility of the monthly average commodity housing sales price index in the last six months, Represents the historical volatility of the futures composite index in the most recent month. Represents the historical volatility of the monthly average futures composite index in the last six months. represents the loan misappropriation index. In this embodiment, is 100, is 200, is 300, is 1, is 0, is 2, is 0, is 1, is 3, is 0.2, is 0.3, is 0.1, is 0.2, is 0.4, is 0.5, so is 25, is 30, is 405, It is 153.

[0121] Embodiment 4: This embodiment is based on embodiment 3, and the behavior analysis module constructs a fusion integrated neural network model, which specifically includes the following steps:

[0122] Step B1: Constructing behavior recognition dataset;

[0123] Step B2: construct a fusion integrated neural network model;

[0124] Step B3: Use the behavior recognition dataset to train the fusion ensemble neural network model.

[0125] Example 5, see Figure 4 and Figure 5 This embodiment is based on the fourth embodiment, and the step B1 specifically includes the following steps:

[0126] Step B11: collecting past mouse track and typing rhythm data. In this embodiment, the pynput library is used to collect past mouse track and typing rhythm data.

[0127] Step B12: extracting a mouse coordinate time series, a mouse event time series and a key time series from the past mouse trajectory and typing rhythm data, wherein the mouse coordinate time series records the coordinates of the mouse on the screen along the time axis, the mouse event time series records the mouse click events along the time axis, and the key time series records which keyboard keys are pressed at each moment along the time axis;

[0128] Step B13: superimpose the mouse coordinate time series and the mouse event time series to generate a multi-channel mouse time series, aggregate the multi-channel mouse time series and the key time series to generate a positive sample of the behavior recognition data set, and add a true or false mark to the positive sample of the behavior recognition data set, and the true or false mark value is 1;

[0129] Step B14: randomly generate multi-channel mouse time series and key time series, and aggregate them into negative samples of the behavior recognition data set. The true and false label value of the negative samples of the behavior recognition data set is 0;

[0130] Step B15: Mix the negative samples of the behavior recognition data set with the positive samples of the behavior recognition data set to generate a behavior recognition data set.

[0131] Example 6, see Figure 2 This embodiment is based on the fifth embodiment, and the step B2 specifically includes the following steps:

[0132] Step B21: construct a concatenation layer, a fifth fully connected layer, and a sixth fully connected layer and connect them in sequence to construct a fusion feature branch, which performs feature fusion on its input and obtains a first robot probability;

[0133] Step B22: construct a one-dimensional convolutional layer, a Flatten layer, a first fully connected layer, and a second fully connected layer, and connect them in sequence, connect the Flatten layer to the concatenation layer, and construct a mouse trajectory feature branch. The mouse trajectory feature branch extracts the comprehensive mouse features in its input and shares them to the fusion feature branch, and finally obtains the second robot probability;

[0134] Step B23: constructing an LSTM layer, a first fully connected layer, and a second fully connected layer, and connecting them in sequence, connecting the LSTM layer to the concatenation layer, constructing a typing rhythm feature branch, wherein the typing rhythm feature branch extracts the typing rhythm feature in its input and shares it to the fusion feature branch, and finally derives the third robot probability;

[0135] Step B24: Specify the one-dimensional convolution layer and the LSTM layer as the input layer, specify the second fully connected layer, the sixth fully connected layer, and the fourth fully connected layer as the output layer, specify the optimizer as Adam, specify the binary cross entropy as the loss function, and build a fusion integrated neural network model.

[0136] Embodiment 7: This embodiment is based on Embodiment 6, and the step B3 specifically includes the following steps:

[0137] Step B31: Divide the behavior recognition dataset into a training set and a test set in a ratio of 8:2;

[0138] Step B32: Specify the multi-channel mouse time series as the input of the one-dimensional convolutional layer, the key time series as the input of the LSTM layer, the true and false labels as the target outputs of the second fully connected layer, the sixth fully connected layer, and the fourth fully connected layer, and train the fusion integrated neural network model.

[0139] Embodiment 8: This embodiment is based on embodiment 7. The process of using the fusion integrated neural network model to derive the probability that the client is a robot based on the client's mouse trajectory and typing rhythm data specifically includes the following steps:

[0140] Step R1: Generate a multi-channel mouse time series and key time series of the customer according to the customer's mouse trajectory and typing rhythm data;

[0141] Step R2: Input the customer's multi-channel mouse time series and key time series into the fusion integrated neural network model, and average the outputs of the second fully connected layer, the sixth fully connected layer, and the fourth fully connected layer to obtain the probability that the customer is a robot.

[0142] Embodiment 9, based on Embodiment 8, the process of the financial analysis module obtaining the probability of material fraud and the probability of credit default according to the information uploaded by the customer specifically includes the following steps:

[0143] Step F1: extract the community name from the address certificate, query the transaction prices of all real estates in the community in the last month based on the community name, and average them to obtain the average housing price in the last month;

[0144] Step F2: Extract the deposit balance and average monthly income in the most recent year from the bank statement;

[0145] Step F3: Extract the first digit of all amounts in bank statements and credit reports, and create a first digit frequency table;

[0146] Step F4: Calculate the KS value between the first digit frequency table and the Benford distribution as the probability of material fraud;

[0147] Step F5: Calculate the customer's expected repayment ability. The calculation formula of the expected repayment ability is as follows:

[0148] ;

[0149] in, Represents expected repayment capacity, Represents the average housing price in the most recent month. Represents the deposit balance, Represents the average monthly income in the most recent year, Represents the number of months of the loan term. In this embodiment, 500,000 yuan, 100,000 yuan, is 5,000 yuan, is 24, so 725,000 yuan;

[0150] Step F6: Collect past credit records and build a credit multidimensional data warehouse;

[0151] Step F7: constructing an Apriori association rule mining model using a credit multidimensional data warehouse. In this embodiment, Microsoft SQL Server Analysis Services is used to construct the Apriori association rule mining model;

[0152] Step F8: Use the Apriori association rule mining model to calculate the credit default probability and relative default probability.

[0153] Embodiment 10: This embodiment is based on embodiment 9, and step F6 specifically includes the following steps:

[0154] Step F61: Create other personal information tables, bank statement tables, personal credit tables, housing price tables and loan record tables from past credit records;

[0155] Step F62: Create an expected repayment capacity table based on the bank statement and address information table;

[0156] Step F63: Designate other personal information tables as fact tables, other personal information tables, bank flow tables, personal credit tables, address information tables, loan record tables and expected repayment capacity tables as dimension tables, and establish a credit multidimensional data warehouse using a star model.

[0157] Embodiment 11: This embodiment is based on Embodiment 10, and the step F8 specifically includes the following steps:

[0158] Step F81: Add other personal information to the other personal information table, add the average housing price in the latest month to the housing price table, add the bank flow to the housing price table, and add the loan term and loan amount to the loan record table;

[0159] Step F82: Select other personal information tables, housing price tables, housing price tables and loan record tables, and use the Apriori association rule mining model to obtain the customer's relative default probability. In this embodiment, the prediction function of the Apriori association rule mining model is used to make predictions, and the screening function is used to screen the customer's default probability from the prediction results as the relative default probability;

[0160] Step F83: Add the personal credit report to the personal credit table;

[0161] Step F84: Select other personal information tables, personal credit tables and loan record tables and use the Apriori association rule mining model to obtain the customer's credit default probability. In this embodiment, the prediction function of the Apriori association rule mining model is used to make predictions and the customer's default probability is filtered out from the prediction results using the screening function as the credit default probability;

[0162] Example 12, see Figure 3 This embodiment is based on the eleventh embodiment, and the process of the comprehensive analysis module determining whether the application is approved and how to respond specifically includes the following steps:

[0163] Step C1: Setting the loan misappropriation index threshold, fraud probability threshold, and default probability threshold;

[0164] Step C2: If the sum of the probability of material forgery and the probability that the customer is a robot is greater than the threshold of the probability of forgery, the application is rejected and a reply is given, "Hello, since your identity and application materials need further verification, please bring relevant documents to go offline for processing." In this embodiment, the sum of the probability of material forgery and the probability that the customer is a robot is less than the threshold of the probability of forgery;

[0165] Step C3: If the credit default probability is greater than the default probability threshold, the application is rejected and a reply is given, "Hello, after evaluation, your loan application is rejected due to a bad credit record." In this embodiment, the credit default probability is less than the default probability threshold;

[0166] Step C4: If the relative default probability is greater than the default probability threshold, the application is rejected and a reply is given, “Hello, after evaluation, your loan application is rejected because the loan amount exceeds your financial capacity. We recommend that you reconsider the loan amount.” In this embodiment, the relative default probability is less than the default probability threshold.

[0167] Step C5: If the application is not rejected in steps C2 to C4, the application is approved. In this embodiment, the application is approved;

[0168] Step C6: If the loan misappropriation index is greater than the loan misappropriation index threshold, the customer is marked as a fund misappropriation supervision subject and a reply is given: "Hello, according to regulatory requirements, we will monitor the use of loan funds. Please ensure that the loan funds are not used for high-risk investments to avoid violating relevant regulations." In this embodiment, the loan misappropriation index is greater than the loan misappropriation index threshold, and the customer is marked as a fund misappropriation supervision subject.

[0169] Example 13. This example is based on Example 12. This example runs in the Huawei FusionServer 2488 V5 environment, relies on Anaconda3, and uses Keras as the model framework of the fusion integrated neural network model. The Pandas library and the Numpy library are used to complete the construction of the behavior recognition data set.

[0170] The present invention and its implementation are described above, and such description is not restrictive. The drawings show only one implementation of the present invention, and the actual structure is not limited thereto. In short, if ordinary technicians in the field are inspired by the present invention, without departing from the purpose of the invention, they can design a structure and embodiment similar to the technical solution without creativity, which should fall within the protection scope of the present invention.

Claims

1. A financial business risk control management system based on a big data model, characterized by: include: Technical information collection module, emotional information collection module, fund misappropriation risk assessment module, user interaction module, financial analysis module, behavioral analysis module and comprehensive analysis module; The technical information collection module collects investment index history records and inputs them into the fund misappropriation risk assessment module, and the emotional information collection module collects financial-related video information from the streaming media service platform to calculate the stock market investment sentiment index, real estate investment sentiment index and futures investment sentiment index, and inputs them into the fund misappropriation risk assessment module; The user interaction module prompts the customer to upload the information required for the credit application when applying for a loan, and simultaneously records the customer's mouse trajectory and typing rhythm data. The information uploaded by the customer includes bank statements. The user interaction module inputs the bank statements into the fund misappropriation risk assessment module, inputs the information uploaded by the customer into the financial analysis module, and inputs the customer's mouse trajectory and typing rhythm data into the behavior analysis module; The behavior analysis module constructs a fusion integrated neural network model and uses the fusion integrated neural network model to obtain the probability that the customer is a robot based on the customer's mouse trajectory and typing rhythm data and transmits it to the comprehensive analysis module; The financial analysis module obtains the material fraud probability, credit default probability and relative default probability based on the information uploaded by the customer and transmits them to the comprehensive analysis module; The fund misappropriation risk assessment module obtains the loan misappropriation index based on the information provided by the technical information collection module and the emotional information collection module and the bank flow and inputs it into the comprehensive analysis module; The comprehensive analysis module determines whether the application is approved and how to respond; The behavior analysis module constructs a fusion integrated neural network model comprising the following steps: Step B1: Constructing behavior recognition dataset; Step B2: construct a fusion integrated neural network model; Step B3: Use the behavior recognition dataset to train the fusion ensemble neural network model; The step B1 specifically comprises the following steps: Step B11: collecting past mouse trajectory and typing rhythm data; Step B12: extracting the mouse coordinate time series, the mouse event time series and the key time series from the past mouse trajectory and typing rhythm data; Step B13: superimpose the mouse coordinate time series and the mouse event time series to generate a multi-channel mouse time series, aggregate the multi-channel mouse time series and the key time series to generate a positive sample of the behavior recognition data set, and add a true or false mark to the positive sample of the behavior recognition data set, and the true or false mark value is 1; Step B14: randomly generate multi-channel mouse time series and key time series, and summarize them into negative samples of the behavior recognition data set. The true and false label value of the negative samples of the behavior recognition data set is 0; Step B15: Mixing the negative samples of the behavior recognition data set with the positive samples of the behavior recognition data set to generate a behavior recognition data set; The step B2 specifically comprises the following steps: Step B21: constructing a fusion feature branch, which performs feature fusion on its input and outputs a first robot probability; Step B22: constructing a mouse trajectory feature branch, wherein the mouse trajectory feature branch extracts the comprehensive mouse features in its input and shares them with the fusion feature branch, and finally outputs the second robot probability; Step B23: constructing a typing rhythm feature branch, wherein the typing rhythm feature branch extracts the typing rhythm features in its input and shares them with the fusion feature branch, and finally outputs the third robot probability; Step B24: Specify the one-dimensional convolution layer and the LSTM layer as the input layer, specify the second fully connected layer, the sixth fully connected layer, and the fourth fully connected layer as the output layer, specify the optimizer as Adam, specify the binary cross entropy as the loss function, and build a fusion integrated neural network model.

2. According to the financial business risk control management system based on big data model in claim 1, it is characterized by: The process of the emotional information collection module collecting the finance-related video information of the streaming media service platform to calculate the stock market investment sentiment index, the real estate investment sentiment index and the futures investment sentiment index specifically includes the following steps: Step S1: Select videos containing finance tags from all videos on the streaming media service platform, and obtain the playback volume of the videos containing finance tags; Step S2: sort the videos containing the finance tag in the last month from high to low according to the number of views, and construct a collection of finance videos; Step S3: Download the videos in the financial video collection and obtain the release time, number of bullet screens, number of comments, number of favorites, number of shares, number of followers and number of likes of each video in the financial video collection; Step S4: TF-IDF matrixing of text and calculating sentiment value of annotated keywords; Step S5: extracting acoustic features; Step S6: Calculate the video emotion index of each video in the financial video set; Step S7: Classify the videos in the financial video set into a stock market video set, a real estate video set, and a futures video set according to the labels, while keeping their order; Step S8: calculating the investment sentiment index of the stock market video set as the stock market investment sentiment index; Step S9: Calculate the real estate investment sentiment index and the futures investment sentiment index.

3. According to the financial business risk control management system based on big data model in claim 1, it is characterized in that: The step B3 specifically comprises the following steps: Step B31: Divide the behavior recognition dataset into a training set and a test set in a ratio of 8:2; Step B32: Specify the multi-channel mouse time series as the input of the one-dimensional convolutional layer, the key time series as the input of the LSTM layer, the true and false labels as the target outputs of the second fully connected layer, the sixth fully connected layer, and the fourth fully connected layer, and train the fusion integrated neural network model.

Images