A financial comprehensive risk prediction method, system and medium based on LSTM model

Through the comprehensive financial risk prediction method based on the LSTM model, the use of mirror data and cluster analysis technology to adjust the traffic characteristic parameters, the problem of the impact of fluctuation cycles is solved, and the personalized prediction and risk level assessment of the traffic data of financial institutions is realized, and the risk level assessment is improved. prediction accuracy.

CN118378899BActive Publication Date: 2024-11-05ZHEJIANG FINANCIAL COMPREHENSIVE SERVICE PLATFORM MANAGEMENT CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410815586.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-06-24
Publication Date
2024-11-05
Estimated Expiration
2044-06-24

AI Technical Summary

Technical Problem

The existing technology cannot effectively avoid the impact of fluctuation cycles in the network traffic forecast of financial institutions, resulting in a decrease in the accuracy of the prediction results and the inability to personalize risk level warnings for different financial institutions.

Method used

Using the comprehensive financial risk prediction method based on the LSTM model, the feature matrix is ​​generated through mirror data, Kmeans clustering analysis is performed, offset components are determined, log time series is corrected, neural network is trained, decision tree model is established, optimal traffic cycle is obtained, and adjustment is made. Feature parameters, perform reorganization prediction of traffic data, generate dynamic prediction templates, and avoid the impact of fluctuation cycles.

Benefits of technology

It realizes personalized prediction of traffic data of different financial institutions, avoids the impact of fluctuations such as holidays on prediction accuracy, and does not require repeated training of prediction models, which improves the accuracy and pertinence of prediction.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118378899B_ABST
    Figure CN118378899B_ABST
Patent Text Reader

Abstract

The present invention discloses a financial comprehensive risk prediction method, system and medium based on LSTM model, comprising: generating a first weight parameter based on a feature matrix according to mirror data of flow data; splitting the mirror data to obtain a log time series, and obtaining a mirror data distribution diagram of the log time series through Kmeans clustering analysis; training to obtain a first feedback number of a neural network after correcting the log time series; establishing a decision tree model of flow data to obtain an optimal flow cycle; assigning a first weight parameter to the first feedback number to form a feature parameter adjustment amount corresponding to the optimal flow cycle; establishing a mapping between the feature parameter adjustment amount and the optimal flow cycle, dividing and reorganizing the flow data based on indicators requested by different interfaces and corresponding mappings, and predicting based on the reorganized flow data; avoiding the influence of the fluctuation cycle, generating fused data based on different feature parameter adjustment amounts for dynamic prediction.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of financial supervision technology, and in particular to a method, system and medium for predicting comprehensive financial risks based on an LSTM model. Background Art

[0002] With the rapid development of network technology and the popularization of Internet technology, the network has become an indispensable part of people's work and life. Traditional network traffic prediction mainly relies on subjective experience, and predicts and judges traffic based on historical data and personal experience. This type of method has a high probability of misjudgment and cannot make accurate judgments on traffic change trends. The prediction results are significantly different from the actual data. If traffic prediction is based on a large amount of historical data and real-time grasp of the current network status, the results will be more reliable. In recent years, many traffic prediction models have been proposed, including statistical and regression methods, traffic-based methods, and machine learning methods. The currently widely recognized autoregressive integrated moving average model has been applied in the field of network security to predict the status of network systems.

[0003] In the process of digital transformation and improving financial services, financial institutions have an urgent need for government data. As a data bridge between the government and financial institutions, the comprehensive financial service platform needs to strengthen the safe use of data while serving financial institutions. Currently, the commonly used monitoring of network traffic is limited to post-judgment, and the definition of the threshold rules for the call volume of different financial institutions is basically the same, and there is only a single warning prompt, which cannot divide the network traffic of different financial institutions in a targeted manner, and then cannot issue warnings of different levels of risk.

[0004] For example, a method and device for detecting anomaly of industrial control network traffic based on deep learning, which is disclosed in Chinese patent literature and has the announcement number CN113162811A, discloses the following steps: obtaining the data to be detected and historical data of the network traffic and preprocessing them to obtain traffic characteristics; selecting the traffic characteristics corresponding to a section of historical data and inputting them into the LSTM prediction model to predict the traffic characteristics corresponding to the data to be detected; calculating the dynamic threshold for anomaly detection based on the historical data, combining the prediction results with the actual traffic characteristics corresponding to the data to be detected, and detecting whether network traffic anomalies occur in the time period corresponding to the data to be detected; storing the traffic characteristic data and regularly training a new LSTM prediction model. If the new LSTM prediction model is better than the current LSTM prediction model used to predict the traffic characteristics, the current LSTM prediction model used to predict the traffic characteristics is updated. However, this solution cannot solve the impact of fluctuations in bank traffic. Summary of the invention

[0005] In order to solve the problem in the prior art that the fluctuation cycle affects the prediction results, the present invention provides a financial comprehensive risk prediction method, system and medium based on the LSTM model, which avoids the influence of the fluctuation cycle and generates fused data based on different feature parameter adjustment amounts for dynamic prediction.

[0006] In order to achieve the above object, the present invention provides the following technical solutions:

[0007] A financial comprehensive risk prediction method based on an LSTM model, comprising: generating a first weight parameter based on a feature matrix according to mirror data of flow data;

[0008] The mirror data is split to obtain a log time series, and a mirror data distribution diagram of the log time series is obtained by Kmeans cluster analysis;

[0009] Determine an offset component based on the traffic data feature in the mirror data distribution graph, correct the log time series according to the offset component, and then train to obtain a first feedback number of the neural network;

[0010] Establishing a decision tree model of traffic data to obtain an optimal traffic cycle; assigning a first weight parameter to the first feedback number to form a characteristic parameter adjustment amount corresponding to the optimal traffic cycle;

[0011] Establish a mapping between characteristic parameter adjustment and optimal traffic cycle, divide and reorganize traffic data based on indicators and corresponding mappings requested by different interfaces, and make predictions based on the reorganized traffic data. Parameter adjustment is obtained through simulation prediction of mirror data of each financial institution, and the parameter adjustment is fed back based on traffic characteristics of different financial institutions in different time periods to generate a dynamic prediction template used by the comprehensive platform, so as to reorganize and predict the traffic data of different financial institutions respectively, avoiding the impact of fluctuations such as holidays on prediction accuracy, and there is no need to repeatedly train the prediction model.

[0012] Preferably, after acquiring the mirror data, a data transfer matrix of the mirror data is generated through a state transfer equation, a feature matrix is ​​generated according to the feature distribution of the data transfer matrix, and the feature matrix is ​​weighted to obtain a first weight parameter, so as to achieve the influence of the acquired data distribution characteristics on the model training.

[0013] Preferably, in the Kmeans clustering analysis, the mean square error of the elements in the log time series is used for clustering, and the length of the log time series is used as the range parameter of the clustering; multiple second-class cluster points are generated based on the log time series, so as to remove the influence of discrete data in the traffic data on the prediction.

[0014] Preferably, in the mirror data distribution graph, a multi-objective project group of business indicators is established, and a first type of clustering point is established according to the time distribution of the multi-objective project group;

[0015] The data identification accuracy is generated based on the positions of the first type of clustering points and the second type of clustering points, and the log time series of the mirrored data is corrected according to the data identification accuracy, so as to realize the data correction at the business level in the traffic data.

[0016] As a preferred method, the modified log time series is extracted and filled into a training set template for training to obtain multiple first feedback numbers of the neural network; the training set template is a template corresponding to the indicators in the multi-objective project group. Parameter acquisition based on the business level in the traffic data is realized.

[0017] Preferably, after obtaining the interface request, the flow data corresponding to the interface request is divided into period flows of different lengths, and a regression decision tree is fitted according to the period flows to obtain the optimal period length, so as to determine the optimal training data.

[0018] Preferably, during the prediction, the optimal flow period and the corresponding characteristic parameter adjustment amount corresponding to the day to be predicted are determined, a training set corresponding to the day to be predicted is reorganized and generated, and the day to be predicted is predicted based on the training set, so as to realize dynamic reconstruction of the training set.

[0019] Preferably, after the prediction is performed, the degree of fit is determined based on the R-square of the prediction result, and when the degree of fit meets the corresponding threshold requirement, the corresponding characteristic parameter adjustment amount is extracted for optimization, so as to obtain the optimal parameters.

[0020] A system, applicable to a financial comprehensive risk prediction method based on an LSTM model as described in any embodiment, comprises a financial comprehensive platform and a data security platform connected in sequence, wherein the financial comprehensive platform is connected to a plurality of institutional terminals.

[0021] A medium includes a memory, wherein the memory stores a program executable by a processor;

[0022] A processor, when the processor runs the program in the memory, implements a financial comprehensive risk prediction method based on an LSTM model as described in any embodiment.

[0023] The present invention has the following advantages:

[0024] (1) The parameter adjustment amount is obtained through simulation prediction of the mirror data of each financial institution, and the parameter adjustment amount is fed back in combination with the traffic characteristics of different financial institutions in different time periods to generate a dynamic prediction template used by the comprehensive platform, so as to reorganize and predict the traffic data of different financial institutions respectively, avoiding the impact of fluctuations such as holidays on the prediction accuracy. There is no need to repeatedly train the prediction model, so as to avoid the impact of fluctuation cycles and generate fused data based on different characteristic parameter adjustment amounts for dynamic prediction; (2) Feedback optimization of training process parameters and training data is achieved by considering both the business level and the data feature level. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] The drawings described below are merely exemplary. A person skilled in the art can derive other implementation drawings based on the provided drawings without inventive effort.

[0026] Figure 1 Schematic diagram of the method steps in the embodiment.

[0027] Figure 2 Schematic diagram of the LSTM model training process in the embodiment.

[0028] Figure 3 It is a curve chart of the prediction results of Bank A in the embodiment.

[0029] Figure 4 It is a curve chart of the prediction results of Bank B in the embodiment.

[0030] Figure 5 Schematic diagram of the working logic of the system in the embodiment.

[0031] In the figure:

[0032] 1-Financial institution; 2-Financial integrated platform; 3-Data security platform. DETAILED DESCRIPTION

[0033] Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in the field without making any creative work shall fall within the scope of protection of the present invention.

[0034] like Figure 1-Figure 5 As shown, in a preferred embodiment, the present invention discloses a financial comprehensive risk prediction method based on the LSTM model, comprising the following steps: firstly, generating a first weight parameter based on a feature matrix according to the mirror data of the flow data, determining the weight feature generated based on the mirror data by the first weight parameter, and determining the weight influence representing the mirror data feature by the first weight parameter, thereby avoiding the problem of inconvenience in conversion caused by the original data format. The parameter adjustment amount is obtained by simulating and predicting the mirror data of each financial institution.

[0035] The mirror data is split to obtain a log time series, wherein the log time series is obtained by sorting the mirror data according to the time series and splitting the mirror data according to the log classification to obtain subsequences. The log time series is a mirror data series with multiple segments of different timestamps and log contents.

[0036] The mirror data distribution map of the log time series is obtained through Kmeans clustering analysis to achieve the classification of the mirror data. At this time, each clustering point concentrates several log time series with the same characteristics. The data distribution of the mirror data based on the parameter characteristics is determined through the mirror data distribution map, which facilitates the feature extraction of the mirror data.

[0037] Determine the offset component based on the flow data feature in the mirror data distribution map, and unify the flow data feature into the mirror data distribution map. Prior to this, extract the feature of the flow data from the original data of the flow data to obtain the flow data feature. In the mirror data distribution map, determine the offset of the flow data feature and the cluster point as the offset component, which represents the degree of deviation between the directly obtained data feature and the data feature after the mirror data is clustered.

[0038] The first feedback number of the neural network is obtained by training after correcting the log time series according to the offset component; and the parameter adjustment amount is fed back in combination with the flow characteristics of different financial institutions in different time periods.

[0039] A decision tree model of traffic data is established to obtain the optimal traffic cycle; the most accurate traffic time period of traffic data is determined through the decision tree as the optimal traffic cycle, which reduces the impact of accidental data in traffic data, realizes the screening of traffic data, and improves the reliability of data training.

[0040] The first weight parameter is assigned to the first feedback number to form a characteristic parameter adjustment amount corresponding to the optimal flow period; the deviation between the original flow data and the mirror data can be combined to generate adjustment parameters for the filtered flow data, thereby realizing adjustment of the optimal flow period based on different types of parameters.

[0041] A mapping between characteristic parameter adjustment amount and optimal flow period is established to determine the adjustment parameter corresponding to each optimal flow period. When any flow period is obtained as the optimal flow period, the corresponding characteristic parameter adjustment amount can be determined, thereby achieving targeted adjustment for different optimal flow periods.

[0042] The traffic data is divided and reorganized based on the indicators and corresponding mappings of different interface requests to generate a dynamic prediction template used by the comprehensive platform, so as to reorganize and predict the traffic data of different financial institutions respectively.

[0043] Predictions are made based on the reorganized traffic data, avoiding the impact of fluctuations such as holidays on prediction accuracy, and there is no need to repeatedly train the prediction model.

[0044] In other embodiments, in this solution, when making the prediction, the corresponding optimal flow period is determined for different days to be predicted, and the corresponding characteristic parameter adjustment amount is determined after the optimal flow period corresponding to the day to be predicted is determined. For different days to be predicted, the most accurate flow period corresponding to them changes due to characteristic changes. Therefore, the selection of the optimal flow period based on the day to be predicted avoids the problem that the global optimal flow period has different prediction accuracy for days to be predicted with different characteristics. The prediction accuracy of each day to be predicted is improved.

[0045] The training set corresponding to the day to be predicted is reorganized and generated, wherein the training set includes the optimal flow period, the characteristic parameter adjustment amount corresponding to the optimal flow period and the corresponding mapping, so that the training library of each day to be predicted can be independent of each other and do not interfere with each other.

[0046] The to-be-predicted day is predicted according to the training set, thereby realizing dynamic reconstruction of the training set.

[0047] In other embodiments, after the prediction is performed in this solution, the prediction result is tuned and judged to obtain the optimal parameters. This includes judging the degree of fit based on the R square of the prediction result, and determining whether the prediction result is reliable based on the R square fit degree. When the degree of fit meets the corresponding threshold requirements, it means that the accuracy of the training model prediction meets the requirements, and the accuracy of the parameter prediction meets the requirements at this time; then the corresponding feature parameter adjustment amount is extracted for tuning.

[0048] In other embodiments, the present solution includes obtaining interface traffic mirror data Ri, generating a first data transfer matrix [X(Ri)] based on the interface traffic mirror data Ri through a state transfer equation X(k), obtaining an eigenvector matrix Ki of the first data transfer matrix, and performing weighting according to the eigenvector matrix Ki to obtain a first weight parameter [r1].

[0049] The interface traffic mirror data Ri is semantically identified, the interface log information ri is split, and the interface log information is arranged in time series to form multiple log time series Tri. A multi-objective project group P = {pi, pj, pk} about the business scope of each banking institution is established, and the log time series Tri is divided using the Kmeans clustering algorithm according to the multi-objective project group P to obtain the mirror data distribution map of the corresponding banking institution.

[0050] In the Kmeans clustering algorithm, cluster analysis is performed on the feature vectors of the log time series Tri to obtain a plurality of second-class cluster points based on the log time series Tri.

[0051] The mirror data distribution graph is provided with a first type of clustering point formed by a multi-target project group and a second type of clustering point formed by a clustering algorithm. The Euclidean distance between the first type of clustering point and the second type of clustering point is determined, and the average difference of the Euclidean distance between the first type of clustering point and the second type of clustering point is determined. The data recognition accuracy of the data distribution graph is determined based on the average difference. The accuracy of the log time series Tri is judged based on the data recognition accuracy. The distances between the head and tail of the log recognition sequence Tri and the first type of clustering point and the second type of clustering point are judged respectively. When the distance of one of the items reaches the distance threshold Kp, the log time series Tri is pruned along the direction toward the multi-target project group P based on the data recognition accuracy, including pruning according to the offset component of the log time series Tri from the multi-target project group P in two-dimensional space.

[0052] The offset component is the characteristic vector of the log time series in the two-dimensional space, and is the component in the direction of the direction vector of the position of the multi-target project group.

[0053] Establish the first training set and the first test set, extract the log time series Tri from the mirror data distribution diagram and fill it into the first training set and the first test set. According to the first test set and the first training set, the first forgetting ratio Ft1, the first input ratio It1 and the first output ratio Ot1 of the neural network are trained. The first forgetting ratio Ft1, the first input ratio It1 and the first output ratio Ot1 form the first feedback number [X(N)ti] of the neural network, and the first feedback number [X(N)ti] is weighted according to the first weight parameter [r1] to obtain the characteristic parameter adjustment amount.

[0054] In other embodiments, the steps include sending an interface request, obtaining flow data corresponding to the interface request, dividing the flow data by period Li, and obtaining multiple groups of period flows Lri of different lengths. The multiple groups of period flows have overlapping flows of different lengths. A regression decision tree based on the period flow Lri is established, and the period flow Lri is fitted to obtain a fitting function F(Lri), and the loss function is; wherein Lri is a period flow of length i, and F(Lrj) is a fitting value of a period flow of length j=i+1.

[0055] The optimal cycle length Lrp is determined according to the regression decision tree of the cycle flow Lri, and the flow data is divided according to the optimal cycle length Lrp to obtain at least one set of optimal flow cycles. The optimal flow cycle is used as a unit to establish a sample set including at least one unit. The sample set is subjected to flow prediction through the LSTM long short-term memory network model to obtain the predicted flow Yrp.

[0056] In other embodiments, the distance threshold is determined by the covariance of each item in the multi-target item group P={pi, pj, pk} from the cluster center.

[0057] In another embodiment, the present solution discloses a method for predicting risk of comprehensive financial services, including the following steps. The steps of analyzing and predicting traffic using LSTM include data preprocessing, training LSTM model, testing LSTM model, applying LSTM model and network traffic prediction. The LSTM model includes an input layer, an LSTM layer, a fully connected layer and an output layer, and is used for network traffic security prediction.

[0058] Step 1: Data acquisition and preprocessing Bank institutions initiate requests through the API interface. The data security management and monitoring system collects and analyzes the interface logs to the log server in real time through traffic mirroring, and writes the key features of the logs into the database after normalization analysis, such as source IP, destination IP, destination port, timestamp, request method, and request parameters. The request parameters include: API path, API parameters, account number, and token. Return parameters. The daily call volume data of different financial institutions from January 2023 to March 2024 are selected and normalized.

[0059] Step 2: LSTM model training The LSTM model includes an input layer, an LSTM layer, a fully connected layer, and an output layer. The underlying data is divided into two parts according to the time dimension, of which the data from January 2023 to December 2023 is the training set, and the data from January 2024 to March 2024 is the test set. The data is input into the neural network for training, and the Adam optimizer is used to optimize the network parameters of the LSTM in iterative learning. The local optimal position and the global optimal position of the particle are updated according to the size of the fitness value to obtain the optimal LSTM model.

[0060] Step 3: LSTM model evaluation In order to evaluate the accuracy of the LSTM model, the trained model is used to predict the training set data and compare it with the true value to evaluate the model's prediction ability. The present invention selects the R-square indicator to quantitatively evaluate the quality of the model prediction effect. The call volume predicted by the LSTM model is calculated and compared with the actual call volume. The larger the R-square value, the better, which means that the deviation between the true value and the model prediction value is smaller, that is, the better the model prediction ability. Some results are shown in the attached figure.

[0061] The historical request volume data of banking institutions is compared with the deviation value predicted by the corresponding model, and the corresponding monitoring measures are determined according to the deviation comparison results; the calculation formula of the degree of fit is as follows:

[0062] R2 = 1 - (SSR / SST);

[0063] Among them, SSR stands for the Sum of Squares Residuals, which is the sum of squares of the differences between the model predictions and the actual observed values. SST stands for the Total Sum of Squares, which is the sum of squares of the differences between the actual observed values ​​and the mean of the dependent variable.

[0064] Step 4: Predict future call volumes based on the past daily call volumes of financial institutions, and determine whether the traffic is risky by comparing it with the actual call volumes. If the model outputs suspected risk or confirmed risk, appropriate measures need to be taken, including blocking traffic or alerting administrators.

[0065] The LSTM model analysis results are first verified and adjusted through supervised machine learning methods to obtain optimized prediction results, and a monitoring benchmark for the daily number of API interface requests is built based on the prediction results.

[0066] Real-time monitoring of the number of real-time requests in the API interface based on the monitoring benchmark includes: real-time monitoring of the deviation value between the number of real-time requests of a bank and its corresponding benchmark and the preset threshold, and determining the risk level of the request "normal, suspected risk, confirmed risk" based on the comparison result of the deviation value and the threshold, and taking corresponding early warning measures for the bank by automatically running the strategy script at a scheduled time.

[0067] like Figure 3-Figure 4 As shown in the figure, the model has good predictive power in general, and the neural network model has a broad application space in the prediction of time series data. The R-square values ​​of Bank A and Bank B are both greater than 0.7, and the fitting effect is good.

[0068] In another embodiment, Figure 2 As shown, the traffic prediction by LSTM long short-term memory network model includes the following steps: 1. Obtain historical call volume data from the data security management monitoring system database, including platform characteristics and time series characteristic data. And pre-process the historical traffic data set. Eliminate missing values; calculate the mean and standard deviation of the data, and regard values ​​exceeding the mean plus or minus 3 times the standard deviation as outliers, and eliminate them.

[0069] 2. Normalization. For variables such as the number of visits to the interface and the IP address, statistical processing is performed according to the time dimension with days as the minimum unit to generate time series data, and normalization is performed through x-min / max-min and mapped to between 0 and 1.

[0070] 3. Sample reconstruction and division. The data is divided into training set and test set according to the time dimension. The historical time is used as the training set, and the test set is selected from the historical time and real time. The data from January 2023 to December 2023 is the training set, and the data from January 2024 to March 2024 is the test set.

[0071] 4. Model training and testing to obtain the API interface traffic prediction model. Build an LSTM model, use R-squared to judge the degree of model fitting, and use this value as a reference to tune the model parameters. In the fully connected layer, set the number of layers to 1 to 10, the number of hidden units to 64, 128, and 256, the learning rate to 0.01 to 0.1, the sequence length to 3 to 20, and the dropout to 0.2 to 0.5. Use a brute force method to combine each value. When the global optimum or the maximum number of iterations is reached, the set of parameters corresponding to the optimal R-squared value is finally obtained. Finally, the optimal parameters are 2 layers, 64 hidden units, a learning rate of 0.01, a sequence length of 6, an optimizer of adam, and a dropout of 0.5.

[0072] Make predictions based on the test set and the optimal parameters and output the predicted value at this time, and determine the prediction error corresponding to the test set under the optimal parameters.

[0073] In other embodiments, the present solution includes the following steps: first, through the data security management and monitoring system, logs are collected by means of traffic mirroring, and API traffic is collected, parsed, analyzed, and data monitored.

[0074] Then, build the LSTM model, preprocess the data, and train and test the model.

[0075] Repeat real-time monitoring and adjustment, conduct real-time monitoring, statistics and analysis through the data security management monitoring system, input the historical data of interface traffic corresponding to the day to be predicted of a certain bank into the interface traffic prediction model, and output the interface traffic data corresponding to the day to be predicted.

[0076] Finally, the interface traffic corresponding to a bank's forecast day is compared with the interface traffic counted in real time on that day. If the number of API calls exceeds the threshold, different risk levels are determined, and the bank is notified, paid attention to, and processed with different levels of warnings.

[0077] In other embodiments, in the evaluation of the LSTM model, the prediction effect of the quantitative evaluation model can also be evaluated by the mean absolute error evaluation index.

[0078] In other embodiments, the present solution discloses a system applicable to a financial comprehensive risk prediction method based on an LSTM model as described in any embodiment, including an institutional terminal, the institutional terminal is connected to a financial comprehensive platform, and the financial comprehensive platform is connected to a data security platform. The institutional terminal registers the financial institution to the financial comprehensive platform. The data security platform establishes baseline behavior and makes risk judgments on real-time data. The financial comprehensive platform synchronizes the data response of the institutional terminal to the data security platform.

[0079] When using, Figure 5 As shown, the institutional terminal sends an interface request to the financial integrated platform, and the financial integrated platform responds to the service according to the interface request, and synchronizes the data corresponding to the service response to the data security platform. The data security platform identifies risks based on the real-time traffic data synchronized by the financial integrated platform. The risk identification includes statistics based on the interface requests of the financial institution within a fixed period, and determining the baseline behavior based on the statistical results. The baseline behavior represents the interface request behavior when the traffic of the corresponding financial institution is within a safe range.

[0080] The data security platform judges the current traffic data based on the baseline behavior. When the real-time data does not exceed the baseline behavior, it means that there is no abnormality in the corresponding real-time data; when the real-time data exceeds the baseline behavior, it means that the corresponding real-time data hits the risk rules and there is an abnormality, and message push and risk handling are performed.

[0081] After the data security platform obtains real-time traffic data, it makes predictions based on the LSTM model, including extracting platform features and time series feature data, using the long short-term memory network model to predict the call volume of different financial institutions and output the risk level of the network.

[0082] In other embodiments, the present solution generates a feature matrix according to the feature distribution of the data transfer matrix, wherein the feature matrix includes the transformation features of the data transfer matrix, and the feature matrix is ​​weighted to obtain a first weight parameter to achieve the effect of obtaining data distribution features on model training.

[0083] In other embodiments, the scheme clusters the mean square error of each element in the log time series, determines the mean square error of each element relative to all elements, takes the length of the log time series as the range parameter of clustering, determines the cluster center according to the distribution of the mean square error, and sets the range of the cluster center according to the range parameter. Thus, multiple second-class cluster points are generated based on the log time series, and the influence of discrete data in the traffic data on the prediction is removed.

[0084] In other embodiments, the present scheme first establishes a multi-objective project group of business indicators, wherein the business indicators include data volume indicators of respective businesses of banks and financial institutions; establishes a first type of clustering points according to the time distribution of the multi-objective project group; and realizes feature extraction of data volume in the time dimension.

[0085] Then, the data recognition accuracy is generated based on the positions of the first type of clustering points and the second type of clustering points. At this time, the data recognition accuracy is determined according to the Euclidean distance of the two types of clustering points. The log time series of the mirror data is corrected according to the data recognition accuracy. The data correction at the business level in the traffic data is realized.

[0086] In other embodiments, the scheme also includes extracting the corrected log time series, where the log time series is the original data for this training, and the log time series includes multiple groups of different lengths; filling it into the training set template for training to obtain multiple first feedback numbers of the neural network; and realizing parameter acquisition at the business level based on the traffic data. At this time, the first feedback number is the characteristic value obtained by training the log time series. The training set template is the template corresponding to the indicator in the multi-objective project group.

[0087] In other embodiments, after obtaining the interface request, this solution divides each interface request into periodic traffic exclusive to the interface request. This includes dividing the traffic data corresponding to the interface request into periodic traffic of different lengths, and dividing each time according to the characteristics of the interface request or the requirements of the interface request. Then, fitting is performed in a regression decision tree according to the periodic traffic, and screening is performed through the regression decision tree, where each branch is a periodic traffic of different lengths, and each node is whether a feature exists; the optimal periodic length is obtained. This realizes the determination of the optimal training data.

[0088] In other embodiments, the present solution discloses a medium including a memory, wherein the memory stores a program executable by a processor.

[0089] A processor, when the processor runs the program in the memory, implements a financial comprehensive risk prediction method based on an LSTM model as described in any embodiment.

[0090] These modifications or improvements made on the basis of the present invention without departing from the spirit of the present invention all fall within the scope of protection required by the present invention.

Claims

1. A comprehensive financial risk prediction method based on LSTM model, It is characterized in that include: Generate a first weight parameter based on the feature matrix according to the mirror data of the traffic data; The mirror data is split to obtain a log time series, and a mirror data distribution diagram of the log time series is obtained by Kmeans cluster analysis; In the Kmeans clustering analysis, clustering is performed based on the mean square error of elements in the log time series, and the length of the log time series is used as the range parameter of clustering; multiple second-class clustering points are generated based on the log time series; in the mirror data distribution diagram, a multi-objective project group of business indicators is established, and first-class clustering points are established according to the time distribution of the multi-objective project group; data recognition accuracy is generated based on the positions of the first-class clustering points and the second-class clustering points; Extracting features from the original data of the flow data to obtain flow data features, unifying the flow data features into the mirror data distribution map, and determining the flow data features and the offset of the cluster points in the mirror data distribution map as an offset component; After correcting the log time series according to the data recognition accuracy and the offset component, the first feedback number of the neural network is obtained by training; a decision tree model of the flow data is established to obtain the optimal flow period, and a first weight parameter is assigned to the first feedback number to form a characteristic parameter adjustment amount corresponding to the optimal flow period; A mapping between characteristic parameter adjustment amounts and optimal traffic cycles is established, traffic data is divided and reorganized based on the indicators of different interface requests and the corresponding mappings, and predictions are made based on the reorganized traffic data.

2. According to the financial comprehensive risk prediction method based on the LSTM model according to claim 1, It is characterized in that After acquiring the mirror data, a data transfer matrix of the mirror data is generated through a state transfer equation, a feature matrix is ​​generated according to a feature distribution of the data transfer matrix, and the feature matrix is ​​weighted to obtain a first weight parameter.

3. According to the financial comprehensive risk prediction method based on the LSTM model according to claim 1, It is characterized in that It includes extracting the corrected log time series, filling it into the training set template for training, and obtaining multiple first feedback numbers of the neural network; the training set template is a template corresponding to the indicators in the multi-objective project group.

4. A financial comprehensive risk prediction method based on LSTM model according to claim 1 or 2, It is characterized in that After obtaining the interface request, the flow data corresponding to the interface request is divided into period flows of different lengths, and a regression decision tree is fitted according to the period flows to obtain the optimal period length.

5. According to the financial comprehensive risk prediction method based on LSTM model according to claim 1, It is characterized in that During the prediction, the optimal flow period and the corresponding characteristic parameter adjustment amount corresponding to the day to be predicted are determined, a training set corresponding to the day to be predicted is reorganized and generated, and the day to be predicted is predicted based on the training set.

6. A financial comprehensive risk prediction method based on LSTM model according to claim 5, It is characterized in that After the prediction is performed, the degree of fit is determined based on the R-square of the prediction result. When the degree of fit meets the corresponding threshold requirement, the corresponding characteristic parameter adjustment amount is extracted for optimization.

7. A system, applicable to a financial comprehensive risk prediction method based on an LSTM model as described in any one of claims 1 to 6, It is characterized in that It includes a financial comprehensive platform and a data security platform connected in sequence, wherein the financial comprehensive platform is connected to a plurality of institutional terminals.

8. A medium comprising a memory, wherein the memory stores a program executable by a processor; A processor, when the processor runs the program in the memory, implements a financial comprehensive risk prediction method based on an LSTM model as described in any one of claims 1 to 6.