Cloud computing supported guarantee financial risk control model and implementation method thereof

Abstract

The invention relates to the technical field of financial risk control, and discloses a guarantee financial risk control model supported by cloud computing and an implementation method thereof, and the model integrates a data management module, a model construction module and a risk control module. The data management module is responsible for collecting customer data from a multivariate data source, integrating historical guarantee cases as risk tags, and storing the risk tags in the cloud computing platform. The model construction module uses a machine learning algorithm to train a credit scoring model and a repayment capability prediction model based on the collected data to form a risk assessment system; meanwhile, a residual error fitting model is constructed to optimize an evaluation result. And the risk control module performs real-time evaluation on new customer data by applying the models, and generates accurate risk rating and early warning signals. According to the invention, efficient processing of data and quick response of the model are realized through the cloud computing technology, the accuracy and timeliness of guarantee financial risk control are effectively improved, and a powerful decision support tool is provided for financial institutions and guarantee institutions.

Description

Technical Field

[0001] The present invention relates to the technical field of financial risk control, and in particular to a guarantee financial risk control model supported by cloud computing and an implementation method thereof. Background Art

[0002] In the financial field, guarantee business, as an important financial service, builds a bridge of trust between customers and lenders. However, guarantee business is also accompanied by certain risks, especially when customers cannot repay on time, the guarantor needs to bear the corresponding repayment responsibility, which may cause the guarantee institution to face huge economic losses. Therefore, how to effectively evaluate and control the risks in guarantee business has become an urgent problem for financial institutions and guarantee companies.

[0003] Traditional risk assessment methods mainly rely on simple statistical models, which are incapable of processing large-scale, multi-dimensional data. On the one hand, these models often lack systematicity and comprehensiveness in data collection and processing, and cannot fully utilize the advantages of cloud computing platforms to integrate and analyze multi-source data; on the other hand, the accuracy and stability of existing risk assessment models still need to be improved, especially in the face of complex and changing financial environments, the model's predictive ability may be greatly affected.

[0004] In addition, traditional risk assessment models often only focus on the client's financing and investment situation and the client's legal litigation data, while ignoring other factors that may affect the risk. These factors play an increasingly important role in modern financial risk assessment, but traditional models are difficult to effectively capture and analyze them. Summary of the invention

[0005] The purpose of the present invention is to provide a guarantee financial risk control model supported by cloud computing and an implementation method thereof to solve the problems raised in the above background technology.

[0006] To achieve the above-mentioned purpose, the present invention provides the following technical solution: a guarantee financial risk control model supported by cloud computing, the model comprising:

[0007] The data management module is used to collect relevant customer data from multiple data sources and store the collected data in the data warehouse of the cloud computing platform; at the same time, it integrates historical guarantee cases and their results as risk tag data;

[0008] Model building module, used to build guarantee risk assessment model and residual fitting model, where the risk assessment model includes credit scoring model and repayment capacity prediction model, which are used to assess the credit status and future repayment capacity of customers; based on the collected data set and risk label data, the risk assessment model is trained using machine learning algorithm; the residual fitting model is constructed using the prediction residual of the risk assessment model to correct and optimize the risk assessment results;

[0009] The risk control module uses the trained risk assessment model and residual fitting model to conduct real-time risk assessment on new customer data and generate risk ratings and early warning signals.

[0010] Preferably, the model building module uses a logistic regression (LR) algorithm to build a credit scoring model, and outputs the customer's credit score through feature selection, model training and parameter tuning to measure the customer's probability of default.

[0011] Preferably, the steps of constructing the credit scoring model include:

[0012] From the data set provided by the data management module, based on business understanding and statistical analysis, select characteristic variables related to credit assessment, including the customer's historical borrowing record, credit history length, debt ratio, repayment record and frequency or asset ownership;

[0013] Preprocess the selected feature variables, including missing value processing, outlier processing, data type conversion and feature standardization, and perform feature crossover and feature conversion on the preprocessed data to construct a feature set for model training;

[0014] Use the processed feature set and the corresponding risk label data, where the risk label data indicates whether the customer defaults, to build a logistic regression model; the formula of the logistic regression model is:

[0015]

[0016] Among them, Y is the risk label, Y = 1 means default, Y = 0 means no default; X1, X2, ..., X n are the characteristic variables in the feature set; β0,β1,β2,...,β n is the model parameter, which is estimated through training data. The gradient descent optimization algorithm is used to iteratively solve the model parameters so that the probability distribution predicted by the model is close to the actually observed risk label distribution. The formula for solving the model parameters by the gradient descent optimization algorithm is:

[0017]

[0018] Among them, θ represents the model parameters, including all weights and biases that the model needs to learn; α represents the learning rate, which is used to control the step size of parameter update; represents the gradient of the loss function J(θ) with respect to the parameter θ, which is a vector pointing to the direction in which the loss function grows fastest; := represents the assignment operation, that is, updating the value of the parameter θ;

[0019] The regularization parameters of the model are adjusted through grid search and cross-validation methods to select the optimal model.

[0020] Preferably, the model building module uses a long short-term memory network (LSTM) algorithm to build a repayment ability prediction model, and predicts the customer's future repayment ability by analyzing the customer's financing and investment conditions and the customer's judicial litigation data.

[0021] Preferably, the training step of the repayment ability prediction model includes:

[0022] Obtain the client's financing and investment information and legal litigation data from the data management module, including transaction amount, transaction time, transaction type, repayment amount, repayment time and overdue status; clean and pre-process the acquired data, including removing invalid data, filling missing values ​​and correcting erroneous data;

[0023] The processed data is organized in a time series format to form the customer's transaction and repayment time series data, where the data at each time point contains the transaction or repayment related information at that time point;

[0024] Define the LSTM network structure, including input layer, hidden layer and output layer; the input layer receives serialized transaction and repayment time series data and converts the data into a format acceptable to the model; the hidden layer contains at least one LSTM unit layer, each LSTM unit has a memory function, which is used to capture the long-term dependencies in the time series data; the output layer predicts the customer's future repayment ability based on the output of the hidden layer, and the output is the predicted value of the repayment ability or the classification label;

[0025] Use serialized transaction and repayment time series data and corresponding repayment capacity labels to train the LSTM model; use the back-propagation algorithm to update the model parameters and minimize the loss function between the predicted value and the true value;

[0026] Evaluate the prediction accuracy and stability of the model on the validation set, and adjust the network structure or hyperparameters to optimize performance.

[0027] Preferably, the step of constructing the residual fitting model by the model building module includes:

[0028] Step 1: Residual calculation: Calculate the residual between the actual prediction value and the true risk label of the credit scoring model and the repayment ability prediction model respectively;

[0029] Step 2: Residual feature engineering: Combine the calculated residual with the original features to form a new residual feature set;

[0030] Step 3, model training: Use the support vector regression SVR algorithm to train the residual feature set and build a residual fitting model; the SVR algorithm is optimized by the following formula:

[0031]

[0032] Among them, w is the weight vector, b is the bias term, C is the regularization parameter, δ i is a slack variable used to measure the degree to which the sample point deviates from the target value;

[0033] Step 4: Model application: The output of the residual fitting model is used as a correction term and added to the prediction results of the original risk assessment model to obtain the corrected risk assessment results.

[0034] Preferably, the method for processing missing values ​​of the selected characteristic variables includes:

[0035] For numerical features, the median of the non-missing values ​​of the feature is used to fill the missing values;

[0036] For categorical features, the missing values ​​are filled with the mode of the non-missing values ​​of the feature;

[0037] For missing values ​​in time series data, the average value of the adjacent values ​​is used to fill them;

[0038] For features with obvious business meanings and few missing values, they are manually filled in according to business logic.

[0039] Preferably, the method for processing outliers on the selected characteristic variables includes:

[0040] For numerical features, the 3σ principle is used to identify and process outliers; specifically, the mean and standard deviation of the feature are calculated, and values ​​exceeding the mean ± 3 times the standard deviation are considered outliers.

[0041] Preferably, the method for converting the data type of the screened feature variables includes:

[0042] For date and time data, convert it to specific numeric features, including converting the date to the number of days or months from a specific date.

[0043] Preferably, a method for implementing a guarantee financial risk control model supported by cloud computing comprises the following steps:

[0044] a) Collect relevant customer data from multiple data sources through the data management module and store the data in the data warehouse of the cloud computing platform; at the same time, integrate historical guarantee cases and their results as risk tag data;

[0045] b) In the model building module, based on the collected data set and risk label data, a machine learning algorithm is used to train a risk assessment model, wherein the risk assessment model includes a credit scoring model and a repayment ability prediction model, which are used to assess the customer's credit status and future repayment ability;

[0046] c) using the risk assessment model to predict part of the data set, calculate the prediction residuals, and construct a residual fitting model based on the prediction residuals to correct and optimize the risk assessment results;

[0047] d) Through the risk control module, the trained risk assessment model and residual fitting model are applied to conduct real-time risk assessment on new customer data and generate risk ratings and early warning signals.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] Through the data management module of the cloud computing platform, the present invention can systematically collect, store and efficiently process customer-related data from multiple data sources. The powerful computing and storage capabilities of cloud computing enable the model to easily cope with large-scale, multi-dimensional data sets, providing a solid data foundation for comprehensive and in-depth analysis of customer risks.

[0050] The present invention constructs a risk assessment system that includes a credit scoring model and a repayment ability prediction model, and innovatively introduces a residual fitting model to correct and optimize the risk assessment results. This multi-level and refined assessment method not only takes into account traditional credit scores and repayment abilities, but also captures other subtle differences that may affect risks through the residual fitting model, thereby significantly improving the accuracy and stability of risk assessment. Unlike traditional risk assessment models that only focus on customer credit scores and repayment abilities, the present invention makes full use of the data integration capabilities of the cloud computing platform and comprehensively considers multi-dimensional risk factors such as customer behavior patterns and social relationships. This comprehensive risk assessment method can more accurately reflect the customer's true risk status and provide a more reliable basis for risk control for guarantee institutions. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 This is a working principle diagram of the guarantee financial risk control model supported by cloud computing according to the present invention;

[0052] Figure 2 A flowchart for building a repayment capacity prediction model;

[0053] Figure 3 Graph of the training steps for fitting a model to the residuals. DETAILED DESCRIPTION

[0054] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0055] See also Figure 1-3 The present invention provides a technical solution: a guarantee financial risk control model supported by cloud computing, the model comprising:

[0056] Data management module: The data management module is responsible for collecting comprehensive customer data from multiple data sources. This data includes but is not limited to the customer's basic information, financial status, historical loan records, credit history, repayment records and frequency, and asset ownership. After data collection is completed, the data is stored in the data warehouse of the cloud computing platform to ensure data security, scalability, and efficient access.

[0057] The data management module is also responsible for integrating historical guarantee cases and their results as risk label data. These historical cases include successful guarantee cases and failed guarantee cases, which provide supervision information for model training, help the model learn and identify potential risk factors.

[0058] Model building module: The model building module is the core of the model and is responsible for building the guarantee risk assessment model and the residual fitting model.

[0059] Risk assessment model: The risk assessment model includes a credit scoring model and a repayment capacity prediction model. The credit scoring model assesses the customer's credit status based on factors such as the customer's credit history and financial status; the repayment capacity prediction model predicts the customer's future repayment capacity based on factors such as the customer's liabilities, financing status, and investment status. Both models are trained using machine learning algorithms to automatically learn and extract key features from data, thereby achieving accurate assessment of customer risk.

[0060] Residual fitting model: To further improve the accuracy of risk assessment, the model building module also uses the prediction residuals of the risk assessment model to build a residual fitting model. The residual fitting model corrects and optimizes the risk assessment results by capturing the residual information that is not fully explained in the risk assessment model. This helps to reduce errors in model predictions and improve the accuracy of risk assessment.

[0061] Risk control module: The risk control module is the application layer of the model, responsible for applying the trained risk assessment model and residual fitting model to conduct real-time risk assessment on new customer data. When new customer data is input into the model, the risk control module first uses the risk assessment model to conduct a preliminary assessment of the customer's credit status and repayment ability; then, the residual fitting model is used to calibrate and optimize the preliminary assessment results; finally, risk ratings and early warning signals are generated based on the calibrated assessment results to provide decision support for guarantee institutions.

[0062] The present invention will be further described below in conjunction with Examples 1 to 4:

[0063] Embodiment 1:

[0064] This embodiment is used to describe in detail the implementation of a credit scoring model based on a logistic regression algorithm.

[0065] The specific steps include:

[0066] Step 1: Obtain the original data set of the customer from the data management module. The data set contains data in multiple dimensions such as the customer's basic information, financial information, historical loan records, etc.

[0067] Based on business understanding and statistical analysis, we screen out characteristic variables that are highly relevant to credit assessment. For example, the number of historical loans (continuous variable), the length of credit history (in months, continuous variable), the debt ratio (continuous variable, total debt / income), and judicial litigation data are used as characteristic variables.

[0068] Preprocess the selected feature variables. For missing values, fill or delete them according to the variable type and business logic; for outliers, correct or remove them; for categorical variables, perform one-hot encoding (One-HotEncoding) or label encoding (Label Encoding); for continuous variables, standardize or normalize them according to the distribution.

[0069] Feature construction and selection

[0070] Based on business experience and data analysis results, new feature variables are constructed. For example, the proportion of a customer's monthly repayment amount to monthly income can be calculated as a repayment pressure indicator; or based on historical loan records, the customer's average loan amount, maximum overdue days, etc. can be calculated.

[0071] Feature selection methods (such as IV value, WOE analysis, correlation analysis, etc.) are used to further screen feature variables to ensure that the selected features contribute significantly to model prediction and that there is no serious multicollinearity between the features.

[0072] Step 2: Divide the dataset after preprocessing and feature selection into a training set and a validation set, usually in a ratio of 7:3 or 8:2. The training set is used for model training, and the validation set is used for model evaluation and parameter tuning.

[0073] Logistic regression model construction

[0074] Use the training set data to build a logistic regression model. The formula of the logistic regression model is:

[0075]

[0076] Among them, Y is the risk label, Y = 1 means default, Y = 0 means no default; X1, X2, ..., X n are the characteristic variables in the feature set; β0,β1,β2,...,β n is the model parameter, which is estimated through training data. The gradient descent optimization algorithm is used to iteratively solve the model parameters so that the probability distribution predicted by the model is close to the actually observed risk label distribution. The formula for solving the model parameters by the gradient descent optimization algorithm is:

[0077]

[0078] Among them, θ represents the model parameters, including all weights and biases that the model needs to learn; α represents the learning rate, which is used to control the step size of parameter update; It represents the gradient of the loss function J(θ) with respect to the parameter θ, which is a vector pointing to the direction in which the loss function grows fastest; := represents the assignment operation, that is, updating the value of the parameter θ.

[0079] Step 3: Use the grid search method to set the range of L1 regularization parameters (for feature selection) and L2 regularization parameters (for preventing overfitting). For example, the L1 regularization parameter can be set to [0.01, 0.1, 1, 10], and the L2 regularization parameter can be set to [0.001, 0.01, 0.1, 1].

[0080] Combined with the 5-fold cross-validation method, the training set data is divided into 5 subsets, and the model is trained and validated for each subset. On each subset, different regularization parameter combinations are used to train the model, and the performance indicators of the model on the validation set (such as AUC value, accuracy, etc.) are calculated.

[0081] According to the cross-validation results, the regularization parameter combination with the best performance index is selected as the parameters of the final model. For example, the parameter combination that maximizes the AUC value is selected. The logistic regression model is retrained using the optimal parameter combination to obtain an optimized credit scoring model.

[0082] Step 4: Deploy the optimized credit scoring model to the risk management system to make real-time predictions on new customer data. Input the customer's characteristic variable values, and the model outputs a credit score, which is used to measure the customer's probability of default. According to business needs and risk strategies, set different credit score thresholds to classify and make decisions on customers. For example, customers with credit scores above a certain threshold are considered high-risk customers, and more stringent review measures or loan services are refused; customers with credit scores below a certain threshold are considered low-risk customers, and more favorable loan terms or increased loan amounts are given.

[0083] Embodiment 2:

[0084] The present invention constructs a repayment ability prediction model based on a long short-term memory network (LSTM) for financial institutions to predict the repayment ability of their customers. The institution collects and cleans the financing and investment conditions of customers and the legal litigation data of customers. Then, the data is serialized according to the above method, and an LSTM network containing two layers of LSTM units is constructed. The model is trained with training set data, and the performance of the model is evaluated on the validation set. After multiple hyperparameter tuning and model optimization, a repayment ability prediction model with good performance is finally obtained. The model can predict the future repayment ability of customers based on their historical transactions and repayment records, and provide support for the credit decision-making of financial institutions.

[0085] The following is a detailed description of the implementation method with reference to specific examples:

[0086] Step 1: Data preparation and serialization:

[0087] Data collection: Extract the client’s financing and investment information and legal litigation data from the financial institution’s data warehouse. This data includes but is not limited to the amount, time, and type of each transaction, as well as the amount, time, and whether the repayment is overdue.

[0088] Data cleaning: Preprocess the extracted data, including removing duplicate records, filling missing values ​​(such as using mean, median or interpolation), correcting erroneous data (such as unreasonable transaction amounts or timestamps), etc.

[0089] Data serialization: Organize the cleaned data in a time series format. For example, for each customer, all of their transaction and repayment records are arranged in chronological order to form a time series. Each record can be represented as a vector, containing transaction or repayment related information, such as transaction amount, transaction type (represented by one-hot encoding), repayment amount (if applicable), whether it is overdue (binary representation), etc.

[0090] Step 2: Model construction:

[0091] Design LSTM network structure:

[0092] Input layer: Receive serialized transaction and repayment time series data. Assume that the input vector of each time step is x t , whose dimension is the number of features of transaction and repayment records.

[0093] Hidden layer: contains one or more layers of LSTM units. The number of LSTM units in each layer (i.e. the number of hidden layer units) is determined according to the complexity of the problem and the size of the data. The calculation formula for LSTM units is as follows:

[0094] Forget gate: f t =σ(W f ·[h t-1 ,x t ]+b f )

[0095] Input gate: i t =σ(W i ·[h t-1 ,x t ]+b i )

[0096] Output gate: o t =σ(W o ·[h t-1 ,x t ]+b o )

[0097] Among them, σ is the sigmoid activation function, W f ,W i ,W o is the weight matrix, b f ,b i ,b o is the bias term.

[0098] Output layer: Based on the output of the hidden layer, predict the customer's future repayment ability. The output can be a continuous value (representing a score of repayment ability) or a classification label (such as "high repayment ability", "medium repayment ability", "low repayment ability"). For classification problems, the output layer usually uses a softmax activation function.

[0099] Step 3: Model training:

[0100] Initialize the weights and biases of the LSTM network randomly or use the parameters of the pre-trained model for initialization. Input the serialized transaction and repayment time series data into the LSTM model, calculate the gradient of the loss function through the back-propagation algorithm, and use the optimization algorithm to update the model parameters. During the training process, strategies such as early stopping and learning rate decay can be used to prevent overfitting and improve model performance. During the training process, methods such as dropout or L2 regularization can be used to reduce the complexity of the model and improve the generalization ability of the model.

[0101] Step 4: Performance evaluation and optimization:

[0102] The dataset is divided into training set, validation set and test set. The training set is used to train the model, the validation set is used to evaluate the model performance and adjust the hyperparameters, and the test set is used to finally evaluate the generalization ability of the model.

[0103] Through grid search, random search or Bayesian optimization, adjust the hyperparameters of the LSTM network (such as the number of hidden layer units, number of layers, dropout rate, learning rate, etc.) to find the model with the best performance. According to the evaluation results, adjust the structure or hyperparameters of the LSTM network, such as increasing or decreasing the number of hidden layer units, adjusting the learning rate, increasing the regularization strength, etc., to further improve the performance of the model.

[0104] Embodiment 3:

[0105] The following will combine specific examples to explain in detail how to implement the residual fitting model construction method:

[0106] Step 1: Residual calculation:

[0107] The credit scoring model predicts based on the customer's historical borrowing data and credit record, and outputs a predicted value representing the probability of default. The repayment ability prediction model predicts based on the customer's transaction record, repayment history, and other features, and outputs a predicted value representing the repayment ability.

[0108] For each sample, the actual prediction values ​​of the credit scoring model and the repayment ability prediction model, as well as the corresponding real risk labels (such as whether it is a default, repayment ability level, etc.) are recorded respectively.

[0109] The residual is calculated using the following formula:

[0110] Residual = True risk label - Prediction value

[0111] For example, if the true risk label of a sample is default (represented by 1) and the predicted value of the credit scoring model is 0.7, then the residual is 1-0.7=0.3.

[0112] Step 2: Residual feature engineering:

[0113] Get the original feature set, which includes all the features used for the credit scoring model and repayment ability prediction model, such as the customer's historical borrowing record and transaction frequency.

[0114] The calculated residual is combined with the original feature set to form a new residual feature set. Specifically, the residual can be added to the original feature set as a new feature, or the residual can be combined with the original feature to generate a new feature. The new residual feature set is preprocessed, such as standardization and normalization, to ensure that the dimensions between the features are consistent and improve the training effect of the model.

[0115] Step 3: Model training

[0116] The support vector regression (SVR) algorithm is selected as the training algorithm for the residual fitting model. The SVR algorithm fits the data by finding an optimal hyperplane that makes the prediction error within a certain range, and has good generalization ability and robustness. The SVR algorithm is optimized by the following formula:

[0117]

[0118] Among them, w is the weight vector, b is the bias term, C is the regularization parameter, δ i It is a slack variable used to measure the degree to which the sample point deviates from the target value. The preprocessed residual feature set is used as the input of the SVR algorithm, and the residual between the true risk label and the predicted value is used as the output of the SVR algorithm for model training.

[0119] During the training process, the optimal regularization parameter C, kernel function type and its parameters are selected through cross-validation and other methods to obtain the SVR model with the best performance. After the training is completed, the parameters of the SVR model are obtained, including the weight vector, bias term, etc. These parameters constitute the residual fitting model.

[0120] Step 4: Model application:

[0121] When a new loan sample needs to be risk assessed, it is first input into the original risk assessment model (i.e., credit scoring model and repayment capacity prediction model) to obtain the original prediction results. At the same time, the new loan sample is input into the residual fitting model, and the residual correction term is calculated based on the parameters of the SVR model.

[0122] The residual correction term is added to the original prediction result to obtain the corrected risk assessment result. For example, if the original prediction result is 0.7 and the residual correction term is 0.05, the corrected risk assessment result is 0.75. Based on the corrected risk assessment result, the customer's credit status and repayment ability can be judged more accurately, providing strong support for loan approval, risk control and other decisions.

[0123] Embodiment 4:

[0124] The following describes in detail the implementation methods of missing value processing, outlier processing and data type conversion of feature variables with specific examples.

[0125] Missing value processing of numerical features:

[0126] For numerical features, such as the customer's "annual income" feature, if there are missing values, the median of the non-missing values ​​of the feature is used for filling. The specific steps are as follows:

[0127] ①Count the values ​​of all non-missing values ​​in the "annual income" feature and calculate their median.

[0128] ② Traverse each record in the data set, and for records with missing values ​​for the "annual income" feature, fill their "annual income" values ​​with the calculated median.

[0129] For example, if the median of non-missing values ​​of the "annual income" feature in a data set is 1,000,000 yuan, then for all records with missing "annual income", their "annual income" values ​​will be filled with 1,000,000 yuan.

[0130] Missing value processing of categorical features:

[0131] For categorical features, such as customer financing features, if there are missing values, the mode of the non-missing values ​​of the feature is used to fill in the missing values. The specific steps are as follows:

[0132] ① Count the frequency of occurrence of all non-missing values ​​in the "financing" feature, and find the value with the highest frequency, that is, the mode.

[0133] ② Traverse each record in the data set, and for records with missing values ​​for the "financing" feature, fill their "financing" values ​​with the found mode.

[0134] For example, if the mode of the "financing" feature in a data set is "Medium", then for all records with missing "financing", their "financing" values ​​will be filled with "Medium".

[0135] Handling missing values ​​in time series data:

[0136] For missing values ​​in time series data, such as if the customer's "monthly repayment amount" feature is missing in a certain month, the average value of the preceding and following months is used to fill in the missing values. The specific steps are as follows:

[0137] ① Locate the month where the missing value is located, and obtain the "monthly repayment amount" values ​​of the preceding and following months.

[0138] ② Calculate the average of the "monthly repayment amount" of the adjacent months.

[0139] ③Fill the calculated mean value into the position where the missing value is located.

[0140] Missing value handling for features with obvious business meaning and few missing values:

[0141] For features with obvious business meaning and few missing values, such as the customer's "financing status" feature, if there are missing values, they are manually filled according to business logic. The specific steps are as follows:

[0142] ① Analyze the business meaning of the “financing status” feature and the possible reasons for missing values.

[0143] ② Develop manual filling rules based on business logic and the distribution of missing values. For example, if the missing values ​​are mainly concentrated among young customers.

[0144] ③Manually fill in missing values ​​according to the established rules.

[0145] Outlier handling:

[0146] For outlier processing of numerical features, the 3σ principle is mainly used for identification and processing. Taking the customer's "financing" feature as an example, the specific steps are as follows:

[0147] ① Calculate the mean and standard deviation of the “financing” feature.

[0148] ② According to the 3σ principle, values ​​exceeding the mean ± 3 times the standard deviation are considered outliers.

[0149] For the identified outliers, a variety of processing methods can be adopted, such as replacing with the mean, median, deleting the record, or modifying according to business logic, etc. In this embodiment, the outliers are replaced with the sum (or difference) of the mean and 3 times the standard deviation, whichever is closer to the mean, to reduce the impact of outliers on the model.

[0150] Data type conversion:

[0151] For date and time data, it is necessary to convert it into specific numerical features for model processing. Taking the customer's "application date" feature as an example, the specific steps are as follows:

[0152] ① Determine a base date, such as the start date of the model training set or a specific business date.

[0153] ② Calculate the difference in days or months between the "application date" and the base date in each record, and use it as a new numerical feature.

[0154] The present invention also includes a method for implementing a guarantee financial risk control model supported by cloud computing, comprising the following steps:

[0155] a) Collect relevant customer data from multiple data sources through the data management module and store the data in the data warehouse of the cloud computing platform; at the same time, integrate historical guarantee cases and their results as risk tag data;

[0156] b) In the model building module, based on the collected data set and risk label data, a machine learning algorithm is used to train a risk assessment model, wherein the risk assessment model includes a credit scoring model and a repayment ability prediction model, which are used to assess the financing and investment status of the customer;

[0157] c) using the risk assessment model to predict part of the data set, calculate the prediction residuals, and construct a residual fitting model based on the prediction residuals to correct and optimize the risk assessment results;

[0158] d) Through the risk control module, the trained risk assessment model and residual fitting model are applied to conduct real-time risk assessment on new customer data and generate risk ratings and early warning signals.

[0159] The implementation of this method refers to the above embodiment and will not be described in detail in the specification.

[0160] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0161] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. The guarantee financial risk control model supported by cloud computing is characterized by: The model includes: The data management module is used to collect relevant customer data from multiple data sources and store the collected data in the data warehouse of the cloud computing platform; at the same time, it integrates historical guarantee cases and their results as risk tag data; Model building module, used to build guarantee risk assessment model and residual fitting model, where the risk assessment model includes credit scoring model and repayment capacity prediction model, which are used to assess the credit status and future repayment capacity of customers; based on the collected data set and risk label data, the risk assessment model is trained using machine learning algorithm; the residual fitting model is constructed using the prediction residual of the risk assessment model to correct and optimize the risk assessment results; The risk control module uses the trained risk assessment model and residual fitting model to conduct real-time risk assessment on new customer data and generate risk ratings and early warning signals.

2. The cloud computing-supported guarantee financial risk control model according to claim 1, characterized in that: The model building module uses the logistic regression (LR) algorithm to build a credit scoring model, and outputs the customer's credit score through feature selection, model training and parameter tuning to measure the customer's default probability.

3. The cloud computing supported guarantee financial risk control model according to claim 2 is characterized in that: The steps of constructing the credit scoring model include: From the data set provided by the data management module, based on business understanding and statistical analysis, select characteristic variables related to credit assessment, including the customer's historical borrowing record, credit history length, debt ratio, repayment record and frequency or asset ownership; Preprocess the selected feature variables, including missing value processing, outlier processing, data type conversion and feature standardization, and perform feature crossover and feature conversion on the preprocessed data to construct a feature set for model training; Use the processed feature set and the corresponding risk label data, where the risk label data indicates whether the customer defaults, to build a logistic regression model; the formula of the logistic regression model is: Among them, Y is the risk label, Y = 1 means default, Y = 0 means no default; X1, X2, ..., X n are the characteristic variables in the feature set; β0,β1,β2,...,β n is the model parameter, which is estimated through training data. The gradient descent optimization algorithm is used to iteratively solve the model parameters so that the probability distribution predicted by the model is close to the actually observed risk label distribution. The formula for solving the model parameters by the gradient descent optimization algorithm is: Among them, θ represents the model parameters, including all weights and biases that the model needs to learn; α represents the learning rate, which is used to control the step size of parameter update; represents the gradient of the loss function J(θ) with respect to the parameter θ, which is a vector pointing to the direction in which the loss function grows fastest; := represents the assignment operation, that is, updating the value of the parameter θ; The regularization parameters of the model are adjusted through grid search and cross-validation methods to select the optimal model.

4. The cloud computing-supported guarantee financial risk control model according to claim 1, characterized in that: The model building module adopts the long short-term memory network LSTM algorithm to build a repayment ability prediction model, and predicts the customer's future repayment ability by analyzing the customer's financing and investment situation and the customer's judicial litigation data.

5. The cloud computing supported guarantee financial risk control model according to claim 4 is characterized in that: The training steps of the repayment ability prediction model include: Obtain the client's financing and investment information and legal litigation data from the data management module, including transaction amount, transaction time, transaction type, repayment amount, repayment time and overdue status; clean and pre-process the acquired data, including removing invalid data, filling missing values ​​and correcting erroneous data; The processed data is organized in a time series format to form the customer's transaction and repayment time series data, where the data at each time point contains the transaction or repayment related information at that time point; Define the LSTM network structure, including input layer, hidden layer and output layer; the input layer receives serialized transaction and repayment time series data and converts the data into a format acceptable to the model; the hidden layer contains at least one LSTM unit layer, each LSTM unit has a memory function, which is used to capture the long-term dependencies in the time series data; the output layer predicts the customer's future repayment ability based on the output of the hidden layer, and the output is the predicted value of the repayment ability or the classification label; Use serialized transaction and repayment time series data and corresponding repayment capacity labels to train the LSTM model; use the back-propagation algorithm to update the model parameters and minimize the loss function between the predicted value and the true value; Evaluate the prediction accuracy and stability of the model on the validation set, and adjust the network structure or hyperparameters to optimize performance.

6. The cloud computing supported guarantee financial risk control model according to claim 1, characterized in that: The steps of constructing the residual fitting model by the model construction module include: Step 1: Residual calculation: Calculate the residual between the actual prediction value and the true risk label of the credit scoring model and the repayment ability prediction model respectively; Step 2: Residual feature engineering: Combine the calculated residual with the original features to form a new residual feature set; Step 3, model training: Use the support vector regression SVR algorithm to train the residual feature set and build a residual fitting model; the SVR algorithm is optimized by the following formula: Among them, w is the weight vector, b is the bias term, C is the regularization parameter, δ i is a slack variable used to measure the degree to which the sample point deviates from the target value; Step 4: Model application: The output of the residual fitting model is used as a correction term and added to the prediction results of the original risk assessment model to obtain the corrected risk assessment results.

7. The cloud computing supported guarantee financial risk control model according to claim 3 is characterized in that: Methods for handling missing values ​​of the selected feature variables include: For numerical features, the median of the non-missing values ​​of the feature is used to fill the missing values; For categorical features, the missing values ​​are filled with the mode of the non-missing values ​​of the feature; For missing values ​​in time series data, the average value of the adjacent values ​​is used to fill them; For features with obvious business meanings and few missing values, they are manually filled in according to business logic.

8. The cloud computing supported guarantee financial risk control model according to claim 3 is characterized in that: The methods for handling outliers on the selected feature variables include: For numerical features, the 3σ principle is used to identify and process outliers; specifically, the mean and standard deviation of the feature are calculated, and values ​​exceeding the mean ± 3 times the standard deviation are considered outliers.

9. The cloud computing supported guarantee financial risk control model according to claim 3, characterized in that: Methods for converting the data types of the selected feature variables include: For date and time data, convert it to specific numeric features, including converting the date to the number of days or months from a specific date.

10. A method for implementing a guarantee financial risk control model supported by cloud computing, characterized in that: The following steps are involved: a) Collect relevant customer data from multiple data sources through the data management module and store the data in the data warehouse of the cloud computing platform; at the same time, integrate historical guarantee cases and their results as risk tag data; b) In the model building module, based on the collected data set and risk label data, a machine learning algorithm is used to train a risk assessment model, wherein the risk assessment model includes a credit scoring model and a repayment ability prediction model, which are used to assess the customer's credit status and future repayment ability; c) using the risk assessment model to predict part of the data set, calculate the prediction residuals, and construct a residual fitting model based on the prediction residuals to correct and optimize the risk assessment results; d) Through the risk control module, the trained risk assessment model and residual fitting model are applied to conduct real-time risk assessment on new customer data and generate risk ratings and early warning signals.

Images