Accurate financial risk identification method based on multiple models
Through multi-model fusion and data source synthesis, combined with feature selection and data rebalancing technology, the problem of insufficient accuracy and reliability of financial risk detection in the existing technology is solved, and efficient and accurate identification of company financial risks is achieved.
Patent Information
- Application Number
- CN202410717086.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-04
- Publication Date
- 2024-11-22
AI Technical Summary
The existing technology has room for improvement in the accuracy and reliability of the company's financial risk detection, and it is difficult to identify financial risks efficiently and accurately.
Using a multi-model-based financial risk identification method, the company's financial raw data is obtained and the China Economic and Financial Research Database is integrated to perform data preprocessing, feature selection and data rebalancing, combined with LASSO, SVM, RF, GBDT and LightGBM models, using SMOTE The algorithm improves data balance and generates an optimal classifier for risk detection through fusion model cross-validation.
It realizes efficient and accurate identification of financial risks, and improves the accuracy and reliability of detection.
Smart Images

Figure SMS_1 
Figure SMS_3 
Figure SMS_4
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of financial data processing, and in particular to a method for accurately identifying financial risks based on multiple models. Background Art
[0003] Although many researchers have conducted in-depth research on corporate financial risk detection and proposed various detection methods, the accuracy of these methods still needs to be improved. For example, the analysis of the financial risks of listed companies can only focus on finding a unified risk logic from the anomalies in the financial statements of companies that have experienced financial losses. Classification models are then used to screen out companies with higher financial risks, which can then be investigated and analyzed.
[0004] To this end, further in-depth exploration of financial risk detection and identification methods is needed to improve the accuracy and reliability of identification. Summary of the Invention
[0005] The purpose of the present invention is to provide a method for accurately identifying financial risks based on multiple models, so as to solve the problem in the prior art that it is difficult to identify the financial risks of a company efficiently and accurately.
[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for accurately identifying financial risks based on multiple models, comprising the following steps:
[0007] S1. Obtain the original financial data of the company to be identified and fuse the original data with the China Economic and Financial Research Database to form a fused data set;
[0008] Among them, the company's original financial data includes the company's annual report data and text information;
[0009] S2. Preprocess the data set to eliminate the influence of extreme values on the recognition effect;
[0010] Among them, data preprocessing includes outlier processing, missing value processing and data normalization.
[0011] S3. Select the main factors that affect the identification of financial risks;
[0012] Among them, feature selection based on penalty items and feature selection based on tree models are adopted. The feature selection based on penalty items adopts LASSO model and SVM model to select features, while the feature selection based on tree adopts RF model, GBDT model and LightGBM model to select features. According to the importance of features, a threshold is set to obtain the top-ranked features of each algorithm. Finally, based on the results selected by each model, the voting method is used to screen out the features with the most votes as the main factors affecting financial risk identification.
[0013] S4, improve the SMOTE algorithm data rebalancing;
[0014] Among them, the improved SMOTE algorithm is based on the SMOTE algorithm. By combining the Tomek Link algorithm denoising and COW value calculation, it eliminates the redundant samples and noise labels in the dataset and the intra-class imbalance problem, thereby achieving the purpose of improving the SMOTE algorithm;
[0015] S5. Divide the data after preprocessing, feature selection and rebalancing into training set and test set;
[0016] S6. Use the training set to train the fusion model to obtain the optimal classifier model;
[0017] S7. Send the test set to the optimal classifier model for risk detection, and obtain the detection result, that is, the prediction result.
[0018] Preferably, the preprocessing of the data set includes the following steps:
[0019] S2.1, outlier processing;
[0020] The Z-score method is used to determine outliers, and the formula is:
[0021]
[0022] Where, X i is the i-th data point, is the mean value of the data, S is the standard deviation of the data, Z i is the Z score of the i-th data point; samples with a Z score greater than 3 or less than -3 are considered outliers and replaced; when the value in the column is greater than e i ±3V i , replace it with e i +3V i ; When the value in the column is less than e i -3V i , replace it with e i -3V i , where e i represents the average value of column i, Vi represents the variance of the i-th column;
[0023] S2.2, missing value processing;
[0024] The formula is:
[0025]
[0026] Where Γ is the missing degree, specifically the percentage of missing eigenvalues in the data set to the total data;
[0027] S2.3, data normalization;
[0028] The data normalization method of maximum and minimum standardization is used to scale each indicator to [0, 1] according to a certain ratio. The specific conversion function formula is:
[0029] X'=(xx min ) / (x max -x min )
[0030] Where x min is the minimum value of the data, x max is the maximum value of the data, x max -x min Very poor.
[0031] Preferably, the feature selection based on the penalty term in step S3 specifically includes:
[0032] Penalty term based on LASSO regression; LASSO regression is a linear regression model using L1 regularization. L1 regularization introduces a penalty term to constrain the model to avoid overfitting. The objective function formula of the LASSO regression model is:
[0033] J(β)=∑(y-Xβ) 2 +λβ1
[0034] =∑(y-Xβ) 2 +∑λβ
[0035] =ESS(β)+λl1(β)
[0036] Where λ is the regularization parameter, y is the classification label, X is the feature matrix of the data, β is the independent variable coefficient, ESS(β) is the error sum of squares, and λl1(β) is the penalty term;
[0037] Penalty term based on the SVM model; the loss function of SVM is Hinge loss plus regularization term, the formula is:
[0038]
[0039] Where y i is the classification label of the i-th data, ω is the model parameter, x i is the i-th data, b is the bias, and λ is the regularization parameter.
[0040] Preferably, the tree-based feature selection in step S2 specifically includes:
[0041] The RF-based feature selection method is a feature selection method that obtains the importance score of each feature by calculating the information gain. Each decision tree in RF is composed of a part of samples and a part of features. On each tree, only a part of the features participate in the decision-making. By randomly selecting samples and features, RF can alleviate the overfitting problem. In RF, the importance of features can be obtained by calculating the information gain brought by the features, that is, calculating the information gain brought by the features after being used to split the nodes on each decision tree, and then averaging or summing the information gains on all decision trees to obtain the importance score of each feature. Through RF-based feature selection, features that have important contributions to the prediction of the target variable can be identified;
[0042] The GBDT-based feature selection method uses the importance index of the decision tree in GBDT to evaluate the importance of features. First, a GBDT model is trained using the original feature set to obtain the importance score of each feature in the decision tree. Then, the features are sorted from high to low according to their importance scores. Finally, the K features with the highest scores are selected as the final feature subset based on their importance scores.
[0043] The feature selection method based on LightGBM is an improved feature selection method based on GBDT. LightGBM is based on the histogram-based decision tree algorithm, which processes large-scale data and trains quickly. Its feature selection uses a built-in feature selection function to select features according to the importance score of the feature, and then calculates the feature importance score based on the weighted average of the number of splits of the feature in all trees. By sorting the features according to their importance scores, the top K features can be selected as the final feature subset.
[0044] Preferably, the steps of Tomek Link algorithm denoising are:
[0045] S4.1. Use the Tomek Link algorithm to denoise the dataset, identify sample points in the dataset that are close in distance but have different labels, identify them as noise points, and delete them from the dataset;
[0046] S4.2. For the denoised dataset, use the SMOTE algorithm to perform oversampling to generate new synthetic sample points;
[0047] S4.3. Select the sample points that are connected to the Tomek Link in the dataset as reference sample points.
[0048] Preferably, the COW value is calculated by using the center offset of the k nearest neighbors of an instance as k increases to measure the density of the sample in the feature space. The ISMOTE algorithm uses instances with larger COW values and selects sparse minority class samples for synthesis to increase the sample diversity of the minority class.
[0049] The value of COW in an n-dimensional dataset depends on the number of minority class instances X m The regional center C of its k nearest neighbors k (X m ), C k (X m ) is calculated as:
[0050]
[0051] Where N k (X m ) is X m The set of k nearest neighbors, X q =(X q1 , X q2 ,······,X qm )∈R m . As the value of k increases, σ i (X m ) to measure C k (X m ) is calculated as follows:
[0052] σ i (X m )=d(C i (X m ),C i+1 (X m ))
[0053] Where d represents the distance, i=1, 2,..., k-1. Generally speaking, in sparse areas, σ i (X m ) is usually larger than in dense areas;
[0054] Considering the influence of k-nearest neighbors on a data node, the degree of this influence is described by calculating the absolute error of the migration amount of the k-nearest neighbor center. The COW coefficient is defined by accumulating the absolute error of the change in the position of the k-nearest neighbor center. This coefficient can be used to measure the overall impact of the change in the position of the k-nearest neighbor center. The formula is:
[0055]
[0056] Where, σ i (X m ) is used to measure C k (X m )'s migration degree.
[0057] Preferably, the method for obtaining the optimal classifier model in step S6 is:
[0058] S6.1, Cross-validation of fusion model;
[0059] Divide the dataset into K parts. For each part, use K-1 parts as training data and the remaining part as test data, thus generating K models and averaging their prediction scores for the labels on the test set.
[0060] S6.2. Select ACO_RF, GBDT, XGBoost, and LightGBM as the first-layer learners of the fusion model, feed the training set into the primary classifier of the four model combinations for training, obtain the primary trainer, and generate a secondary training set;
[0061] S6.3. Select LR as the second-layer learner of the fusion model for secondary training to obtain the final prediction model.
[0062] Preferably, in step S6.1, the training set is randomly divided into 5 parts, and 4 parts of the data are randomly selected each time for training, and the remaining 1 part of the data is used for testing. Finally, five models are generated, and then the prediction effects of the test set are averaged to obtain the final prediction result.
[0063] Preferably, the data after preprocessing, feature selection and rebalancing in step S5 is randomly divided into a training set and a test set in a ratio of 7:3.
[0064] Preferably, in step S2, a Transformer model is used to perform feature selection when preprocessing text information.
[0065] Compared with the prior art, the present invention has the following beneficial effects:
[0066] The present invention relates to a multi-model-based accurate financial risk identification method that adopts multi-model fusion and integrates multiple algorithms and data sources to achieve efficient processing of financial data and effective feature extraction, thereby facilitating efficient and accurate identification of financial risks. DETAILED DESCRIPTION
[0067] The following is a clear and complete description of the technical solutions in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0068] The present invention provides a technical solution, a method for accurately identifying financial risks based on multiple models, comprising the following steps:
[0069] S1. Obtain the original financial data of the company to be identified and fuse the original data with the China Economic and Financial Research Database to form a fused data set;
[0070] Among them, the company's original financial data includes the company's annual report data and text information;
[0071] S2. Preprocess the data set to eliminate the influence of extreme values on the recognition effect;
[0072] The preprocessing of the dataset includes the following steps:
[0073] S2.1, outlier processing;
[0074] The Z-score method is used to determine outliers, and the formula is:
[0075]
[0076] Where, X i is the i-th data point, is the mean value of the data, S is the standard deviation of the data, Z i is the Z score of the i-th data point; samples with a Z score greater than 3 or less than -3 are considered outliers and replaced; when the value in the column is greater than e i ±3V i , replace it with e i +3V i ; When the value in the column is less than e i -3V i , replace it with e i -3V i , where e i represents the average value of column i, V i represents the variance of the i-th column;
[0077] S2.2, missing value processing;
[0078] The processing formula is:
[0079]
[0080] Where Γ is the missing degree, specifically the percentage of missing eigenvalues in the data set to the total data;
[0081] S2.3, data normalization;
[0082] The data normalization method of maximum and minimum standardization is used to scale each indicator to [0, 1] according to a certain ratio. The specific conversion function formula is:
[0083] X'=(xx min ) / (x max -x min )
[0084] Where x min is the minimum value of the data, x max is the maximum value of the data, x max -x min Very poor.
[0085] In addition, the Transformer model is used for feature selection when preprocessing text information.
[0086] S3. Select the main factors that affect the identification of financial risks;
[0087] Among them, feature selection based on penalty items and feature selection based on tree models are adopted. The feature selection based on penalty items adopts LASSO model and SVM model to select features, while the feature selection based on tree adopts RF model, GBDT model and LightGBM model to select features. According to the importance of features, a threshold is set to obtain the top-ranked features of each algorithm. Finally, based on the results selected by each model, the voting method is used to screen out the features with the most votes as the main factors affecting financial risk identification.
[0088] Feature selection based on penalty terms specifically includes: penalty terms based on LASSO regression; LASSO regression is a linear regression model using L1 regularization. L1 regularization introduces a penalty term to constrain the model to avoid overfitting. The objective function formula of the LASSO regression model is:
[0089] J(β)=∑(y-Xβ) 2 +λβ1
[0090] =∑(y-Xβ) 2 +∑λβ
[0091] =ESS(β)+λl1(β)
[0092] Where λ is the regularization parameter, y is the classification label, X is the feature matrix of the data, β is the independent variable coefficient, ESS(β) is the error sum of squares, and λl1(β) is the penalty term;
[0093] Penalty term based on the SVM model; the loss function of SVM is Hinge loss plus regularization term, the formula is:
[0094]
[0095] Where y i is the classification label of the i-th data, ω is the model parameter, x i is the i-th data, b is the bias, and λ is the regularization parameter;
[0096] Tree-based feature selection specifically includes: RF-based feature selection method is a feature selection method that obtains the importance score of each feature by calculating information gain. Each decision tree in RF is composed of a part of samples and a part of features. On each tree, only a part of the features participate in the decision. By randomly selecting samples and features, RF can alleviate the overfitting problem. In RF, feature importance can be calculated by calculating the information gain brought by the feature, that is, calculating the information gain brought by the feature after it is used to split the node on each decision tree, and then averaging or summing the information gain on all decision trees to obtain the importance score of each feature. Through RF-based feature selection, those features that have important contributions to the prediction of the target variable can be identified;
[0097] The GBDT-based feature selection method uses the importance index of the decision tree in GBDT to evaluate the importance of features. First, a GBDT model is trained using the original feature set to obtain the importance score of each feature in the decision tree. Then, the features are sorted from high to low according to their importance scores. Finally, the K features with the highest scores are selected as the final feature subset based on their importance scores.
[0098] The feature selection method based on LightGBM is an improved feature selection method based on GBDT. LightGBM is based on the histogram-based decision tree algorithm, which processes large-scale data and trains quickly. Its feature selection uses a built-in feature selection function to select features according to the importance score of the feature, and then calculates the feature importance score based on the weighted average of the number of splits of the feature in all trees. By sorting the features according to their importance scores, the top K features can be selected as the final feature subset.
[0099] S4, improve the SMOTE algorithm data rebalancing;
[0100] Among them, the improved SMOTE algorithm is based on the SMOTE algorithm. By combining the Tomek Link algorithm denoising and COW value calculation, it eliminates the redundant samples and noise labels in the data set and the intra-class imbalance problem to improve the SMOTE algorithm;
[0101] Steps of Tomek Link algorithm denoising:
[0102] S4.1. Use the Tomek Link algorithm to denoise the dataset, identify sample points in the dataset that are close in distance but have different labels, identify them as noise points, and delete them from the dataset;
[0103] S4.2. For the denoised dataset, use the SMOTE algorithm to perform oversampling to generate new synthetic sample points;
[0104] S4.3. Select the sample points that are connected to the Tomek Link in the dataset as reference sample points.
[0105] The COW value calculation uses the center offset of the k nearest neighbors of an instance as k increases to measure the density of the sample in the feature space. The ISMOTE algorithm uses instances with large COW values to select sparse minority class samples for synthesis to increase the sample diversity of the minority class.
[0106] The value of COW in an n-dimensional dataset depends on the number of minority class instances X m The regional center C of its k nearest neighbors k (X m ), C k (X m ) is calculated as:
[0107]
[0108] Where N k (X m ) is X m The set of k nearest neighbors, X q =(X q1 , X q2 ,······,X qm )∈R m . As the value of k increases, σ i (X m ) to measure C k (X m ) is calculated as follows:
[0109] σ i (X m )=d(C i (X m ),Ci+1 (X m ))
[0110] Where d represents the distance, i=1, 2,..., k-1. Generally speaking, in sparse areas, σ i (X m ) is usually larger than in dense areas;
[0111] Considering the influence of k-nearest neighbors on a data node, the degree of this influence is described by calculating the absolute error of the migration amount of the k-nearest neighbor center. The COW coefficient is defined by accumulating the absolute error of the change in the position of the k-nearest neighbor center. This coefficient can be used to measure the overall impact of the change in the position of the k-nearest neighbor center. The formula is:
[0112]
[0113] Where, σ i (X m ) is used to measure C k (X m )'s migration degree.
[0114] S5. After preprocessing, feature selection, and rebalancing, the data is randomly divided into a training set and a test set in a ratio of 7:3;
[0115] S6. Use the training set to train the fusion model to obtain the optimal classifier model;
[0116] Among them, the method to obtain the optimal classifier model is:
[0117] S6.1, Cross-validation of fusion model;
[0118] The training set is randomly divided into 5 parts, and 4 parts of the data are randomly selected for training each time, and the remaining 1 part of the data is used for testing. Finally, five models are generated, and the prediction effects of the test set are averaged to obtain the final prediction result.
[0119] S6.2. Select ACO_RF, GBDT, XGBoost, and LightGBM as the first-layer learners of the fusion model, send the training set to the primary classifier of the four model combinations for training, obtain the primary trainer, and generate a secondary training set.
[0120] Among them, the ACO_RF algorithm is a random forest algorithm based on ant colony optimization, which uses the ant colony optimization algorithm to select decision trees in the integrated forest to maintain the diversity between decision trees and improve the prediction accuracy by compressing the forest size. Through the ACO algorithm, a sub-forest RF containing fewer decision trees is selected. B , which not only significantly improves the prediction accuracy, but also has a lower average similarity, which is better than RF A1, thereby achieving effective optimization of the random forest algorithm.
[0121] S6.3. Select LR as the second-layer learner of the fusion model for secondary training to obtain the final prediction model.
[0122] S7. Send the test set to the optimal classifier model for risk detection, and obtain the detection result, that is, the prediction result.
[0123] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.
[0124] While 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 these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.
Claims
1. A method for accurately identifying financial risks based on multiple models, characterized in that: The following steps are involved: S1. Obtain the original financial data of the company to be identified, and merge the original data with the China Economic and Financial Research Database to form a fused data set; Among them, the company's original financial data includes the company's annual report data and text information; S2, preprocess the data set to eliminate the influence of extreme values on the recognition effect; Among them, data preprocessing includes outlier processing, missing value processing and data normalization. S3. Select the main factors that affect the identification of financial risks; Among them, feature selection based on penalty items and feature selection based on tree models are used. The feature selection based on penalty items uses LASSO model and SVM model to select features, and the feature selection based on tree uses RF model, GBDT model and LightGBM model to select features. According to the importance of features, a threshold is set to obtain the top-ranked features of each algorithm. Finally, based on the results selected by each model, the voting method is used to screen out the features with the most votes as the main factors affecting financial risk identification. S4, Improved SMOTE algorithm data rebalancing; Among them, the improved SMOTE algorithm is based on the SMOTE algorithm. By combining the Tomek Link algorithm denoising and COW value calculation, the redundant samples and noise labels in the data set and the intra-class imbalance problem are eliminated to improve the SMOTE algorithm. S5, divide the data after preprocessing, feature selection and rebalancing into training set and test set; S6. Use the training set to train the fusion model to obtain the optimal classifier model; S7. Send the test set to the optimal classifier model for risk detection, and obtain the detection result, that is, the prediction result.
2. The method for accurately identifying financial risks based on multiple models according to claim 1 is characterized in that: The preprocessing of the dataset includes the following steps: S2.1, outlier processing; The Z-score method is used to determine outliers, and the formula is: Where, X i is the ith data point, is the mean value of the data, S is the standard deviation of the data, and Z i is the Z score of the ith data point; samples with a Z score greater than 3 or less than -3 are considered outliers and replaced; when the value in this column is greater than e i ±3V i , replace it with e i +3V i ; When the value in the column is less than e i -3V i , replace it with e i -3V i , where e i represents the average value of the i-th column, V i represents the variance of the i-th column; S2.2, missing value processing; The formula is: In the formula, Γ is the degree of missingness, specifically the percentage of missing feature values in the data set to all data; S2.3, data normalization; The maximum and minimum standardized data normalization method is used to scale each indicator to [0, 1] according to a certain ratio. The specific conversion function formula is: X'=(x-x min ) / (x max -x min ) In the formula, x min is the minimum value of the data, x max is the maximum value of the data, x max -x min Very poor.
3. The method for accurately identifying financial risks based on multiple models according to claim 1 is characterized in that: The feature selection based on the penalty term in step S2 specifically includes: Penalty term based on LASSO regression; LASSO regression is a linear regression model using L1 regularization. L1 regularization introduces a penalty term to constrain the model to avoid overfitting. The objective function formula of the LASSO regression model is: J(β)=∑(y-Xβ) 2 +λβ1 =∑(y-Xβ) 2 +∑λβ =ESS(β)+λl1(β) In the formula, λ is the regularization parameter, y is the classification label, X is the feature matrix of the data, β is the independent variable coefficient, ESS(β) is the error sum of squares, and λl1(β) is the penalty term; Penalty term based on the SVM model; the loss function of SVM is Hinge loss plus regularization term, the formula is: In the formula, y i is the classification label of the i-th data, ω is the model parameter, x i is the i-th data, b is the bias, and λ is the regularization parameter.
4. The method for accurately identifying financial risks based on multiple models according to claim 1 is characterized in that: The tree-based feature selection in step S3 specifically includes: The RF-based feature selection method is a feature selection method that obtains the importance score of each feature by calculating the information gain. Each decision tree in RF is composed of a part of samples and a part of features. On each tree, only a part of the features are involved in the decision. By randomly selecting samples and features, RF can alleviate the overfitting problem. In RF, the importance of features can be obtained by calculating the information gain brought by the features, that is, calculating the information gain brought by the features after being used to split the nodes on each decision tree, and then averaging or summing the information gains on all decision trees to obtain the importance score of each feature. Through RF-based feature selection, those features that have important contributions to the prediction of the target variable can be identified; The feature selection method based on GBDT is a feature selection method that uses the importance index of the decision tree in GBDT to evaluate the importance of features. First, a GBDT model is trained using the original feature set to obtain the importance score of each feature in the decision tree. Secondly, the features are sorted from high to low according to their scores, and finally, the K features with the highest scores are selected as the final feature subset according to their importance scores. The feature selection method based on LightGBM is an improved feature selection method based on GBDT. LightGBM is based on the histogram-based decision tree algorithm, processes large-scale data and trains quickly. Its feature selection uses a built-in feature selection function to select features according to the importance score of the feature, and then calculates the feature importance score based on the weighted average of the number of splits of the feature in all trees. By sorting the features according to their importance scores, the top K features can be selected as the final feature subset.
5. The method for accurately identifying financial risks based on multiple models according to claim 1, characterized in that: Steps of Tomek Link algorithm denoising: S4.
1. Use Tomek Link algorithm to denoise the data set, identify sample points with similar distance but different labels in the data set, identify them as noise points, and delete them from the data set; S4.2, for the denoised data set, use the SMOTE algorithm to perform oversampling to generate new synthetic sample points; S4.
3. Select the sample points that are connected to the Tomek Link in the data set as reference sample points.
6. The method for accurately identifying financial risks based on multiple models according to claim 1, characterized in that: The COW value calculation uses the center offset of the k nearest neighbors of an instance as k increases to measure the density of the sample in the feature space. The ISMOTE algorithm uses instances with larger COW values and selects sparse minority class samples for synthesis; The value of COW in an n-dimensional dataset depends on the number of minority class instances X m The regional center C of its k nearest neighbors k (X m ), C k (X m ) is calculated as: Where N k (X m ) is X m The set of k nearest neighbors, X q =(X q1 , X q2 ,······,X qm )∈R m , with the increase of k value, using σ i (X m ) to measure C k (X m ) is calculated as follows: σ i (X m )=d(C i (X m ),C i+1 (X m )) Where d represents the distance, i = 1, 2, ..., k-1. Generally speaking, in sparse areas, σ i (X m ) is usually larger than in dense areas; Considering the influence of k nearest neighbors on a data node, the absolute error of the migration amount of the k nearest neighbor center is calculated to describe the degree of this influence. The absolute error of the change of the k nearest neighbor center position is accumulated to define the COW coefficient, which can be used to measure the overall impact of the change of the k nearest neighbor center position. The formula is: In the formula, σ i (X m ) is used to measure C k (X m )'s migration degree.
7. The method for accurately identifying financial risks based on multiple models according to claim 1 is characterized in that: The method for obtaining the optimal classifier model in step S6 is: S6.1, cross-validation of fusion model; Divide the dataset into K parts. For each part, use K-1 parts as training data and the remaining part as test data, thus generating K models and averaging their prediction scores for the labels on the test set. S6.2, select ACO_RF, GBDT, XGBoost and LightGBM as the first-layer learners of the fusion model, send the training set to the primary classifier of the four model combinations for training, obtain the primary trainer, and generate a secondary training set; S6.
3. Select LR as the second-layer learner of the fusion model for secondary training to obtain the final prediction model.
8. The method for accurately identifying financial risks based on multiple models according to claim 7 is characterized by: In step S6.1, the training set is randomly divided into 5 parts, and 4 parts of the data are randomly selected for training each time, and the remaining 1 part of the data is used for testing. Finally, five models are generated, and then the prediction effects of the test set are averaged to obtain the final prediction result.
9. The method for accurately identifying financial risks based on multiple models according to claim 1, characterized in that: In step S5, the data after preprocessing, feature selection and rebalancing are randomly divided into training set and test set in a ratio of 7:
3.
10. The method for accurately identifying financial risks based on multiple models according to claim 1, characterized in that: In step S2, the Transformer model is used to perform feature selection when preprocessing text information.
Citation Information
Cited By
Training method of default risk judgment model and default risk judgment method
CN119809807A
Enterprise financial data risk prediction method based on big data and artificial intelligence
CN119904321A
