Method, device and storage medium for predicting default of bond issuer based on semi-supervised model

Abstract

The present invention relates to the field of computer technology, and discloses a method, device, and storage medium for predicting the default of a bond issuer based on a semi-supervised model, comprising: S1: obtaining the subject data of the bond issuer, wherein the subject data includes news and public opinion information, industrial and commercial information, market evaluation information, and external information, and constructing an indicator system for the credit default risk of the bond issuer through the subject data; S2: constructing underlying features from statistical analysis, business judgment, and derivation to generate underlying factors; S3: establishing a semi-supervised model based on a combination of an unlabeled sample weighting method and a scoring card model; S4: judging and predicting the default risk of the bond issuer based on the semi-supervised model. The modeling method for this time is based on a combination of an unlabeled sample weighting method and a scoring card model, utilizing the risk ranking capability of the XGB classifier trained with positive samples and unlabeled samples to expand the scale of positive samples, using samples with the highest probability of high risk as new positive samples, training the scoring card model, and constructing a semi-supervised model.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular provides a method, device and storage medium for predicting default of a bond issuer based on a semi-supervised model. Background Art

[0002] Traditional methods for predicting defaults of bond-issuing companies mainly use financial data and credit analyst scores to rate companies and derive their default probability. News and public opinion data, being unstructured data, cannot be directly used by computer models and is difficult to use as input for models. Therefore, how to automatically use news and public opinion to establish a model for predicting defaults of bond issuers is a necessary problem that existing technologies need to solve.

[0003] Existing technologies often use financial data and credit researchers to score companies in different dimensions, rate bond-issuing companies, output default probabilities, and manually process public opinion data. A large number of business personnel are involved, and subjective evaluations are input to formulate a large number of early warning rules. This results in insufficient mining and analysis in current technologies, making it difficult to fully implement credit risk assessments, and resulting in low efficiency and reliance on subjective judgments.

[0004] In addition, default by a bond-issuing company is a low-probability event, so when modeling data, there are very few positive samples. How to use existing samples to expand the proportion of positive samples is the key to solving the problem of model distortion. Summary of the Invention

[0005] In order to solve the problems in the existing technology of manual processing of public opinion data, participation of a large number of business personnel and subjective evaluation in formulating early warning rules, the present invention provides a method, device and storage medium for predicting the default of bond issuers based on a semi-supervised model.

[0006] The technical solutions of the present invention are as follows:

[0007] A method for predicting default of a bond issuer based on a semi-supervised model, comprising:

[0008] S1: Obtain the principal data of the bond issuer, including news and public opinion information, industrial and commercial information, market evaluation information, and external information, and construct an indicator system for the credit default risk of the bond issuer based on the principal data;

[0009] S2: Generate underlying factors through statistical analysis, business judgment, and derivation of underlying features;

[0010] S3: Establish a semi-supervised model based on the combination of unlabeled sample weighting method and scorecard model;

[0011] S4: Determine and predict the default risk of the bond issuer based on the semi-supervised model.

[0012] Furthermore, the S1 indicator system rates the bond issuer through basic qualification information, financial and operational information, penalty information, equity pledge information, news and public opinion information, internal and external rating information, and risk-related information.

[0013] Furthermore, the S2 uses statistical indicators such as logarithm, mean, mode and extreme value to mine the potential information of the bond issuer data.

[0014] Furthermore, establishing the semi-supervised model of S3 includes the following steps:

[0015] S21: Optimize the parameters with AUC as the target through grid search and train the XGBoost model to obtain a classifier that can identify whether the sample is labeled.

[0016] S22: Use the calibration classifier to perform probability calibration, and use the output calibration of XGBoost as the probability of the approximate standard;

[0017] S23: Use the calibrated samples and the original negative labels as the modeling target for the subsequent training scorecard;

[0018] S24: Calculate weights using balanced sample weighting;

[0019] S25: Use chi-square binning to convert all features into ordinal categorical variables;

[0020] S26: Analyze the correlation between features and modeling targets, as well as the collinearity between features, to select high-quality features that can be included in the model;

[0021] S27: Manually optimizing feature interpretability;

[0022] S28: Encode the features by weight of evidence and then train the scorecard model;

[0023] S29: Manually review the scoring rules and correct a few rules that do not match the response rate distribution results.

[0024] Furthermore, the scoring card model of S2 is a scoring card model based on logistic regression, which converts the distribution of each feature of the positive sample into evidence weight coding, and then generates a score by combining the evidence weight and β in the regression coefficient. The output data-driven scoring card model reflects the information mined from the data and the model's operational logic, and gives the scoring process of the bond issuer and the proportion of single factor scores.

[0025] Furthermore, the semi-supervised model was tested for discrimination ability through a KS evaluation model, with KS>0.4.

[0026] Furthermore, the range of AUC of S21 is AUC>0.7.

[0027] The present invention further provides a device for predicting the default of a bond issuer based on a semi-supervised model, the device for predicting the default of a bond issuer based on a semi-supervised model comprising:

[0028] Memory, processor, communication bus and a semi-supervised model stored in the memory to predict the default procedure of the bond issuer,

[0029] The communication bus is used to realize the communication connection between the processor and the memory;

[0030] The processor is used to execute the program for predicting the default of the bond issuer based on the semi-supervised model to implement the steps of the method for predicting the default of the bond issuer based on the semi-supervised model as described in any one of the above.

[0031] The present invention also provides a computer-readable storage medium storing executable instructions, wherein the storage medium stores a program for predicting the default of a bond issuer based on a semi-supervised model. When the program for predicting the default of a bond issuer based on a semi-supervised model is executed by a processor, the steps of any one of the above-mentioned methods for predicting the default of a bond issuer based on semi-supervised machine learning are implemented.

[0032] The beneficial effects of the present invention include at least:

[0033] (1) This modeling method is based on a combination of the unlabeled sample weighting method and the scorecard model. The XGB classifier trained with positive and unlabeled samples is used to expand the scale of positive samples to rank risks. The samples with the highest high-risk probability are used as new positive samples to train the scorecard model. Due to the good interpretability and white-box training process of the scorecard model, it is used as the evaluation model for outputting the final results.

[0034] (2) By using the positive sample and unlabeled sample learning methods in semi-supervised learning, the scale of positive samples was expanded and the original severely biased modeling samples were corrected. On the one hand, it faced the possibility of labeled samples in unlabeled samples. On the other hand, it enabled the model to better learn the characteristics of bad samples and reduced the risk of the model fitting more noise due to sample imbalance.

[0035] (3) This method generates models based on machine learning in a data-driven manner, reducing information loss caused by subjective intervention, making risk warnings more objective and more effective in capturing changes in the risk of prior default by the subject. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 This is a flow chart of the present invention for predicting default of a bond issuer based on a semi-supervised model.

[0037] Figure 2 Flowchart of the semi-supervised model of the present invention. DETAILED DESCRIPTION

[0038] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of 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. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0039] Combine Figure 1 and Figure 2 As shown in FIG, a method for predicting default of a bond issuer based on a semi-supervised model includes:

[0040] S1: Obtain the principal data of the bond issuer, including news and public opinion information, industrial and commercial information, market evaluation information, and external information, and construct an indicator system for the credit default risk of the bond issuer based on the principal data;

[0041] S2: Generate underlying factors through statistical analysis, business judgment, and derivation of underlying features;

[0042] S3: Establish a semi-supervised model based on the combination of unlabeled sample weighting method and scorecard model;

[0043] S4: Determine and predict the default risk of the bond issuer based on the semi-supervised model.

[0044] The main goal of the modeling of this invention is to predict entities with a high probability of default through quantitative models and to achieve pre-emptive detection of default risks. The sources of analysis data are mainly corporate business data and news and public opinion data. The analysis object is the entity with public opinion data. It mainly starts from the two perspectives of news and public opinion and basic business information, and explores the potential patterns and connections of the predicted entity in three dimensions: basic qualifications, business changes, and public opinion changes. It enriches the underlying indicators through feature engineering and explores the risk factors associated with default risk. It also establishes a scoring card model based on semi-supervised learning to evaluate the possibility of default risk of the predicted entity.

[0045] To this end, first, we determine the target variable: define the target variable as the bond defaulting entity with prior warning and matching industrial and commercial information, and use it as a severely biased sample;

[0046] Secondly, through feature engineering, we generate underlying factors. Factors generated from industrial and commercial information include basic qualification evaluation, financial report information evaluation, penalty information evaluation, and equity pledge evaluation; factors generated from early warning data include the time, exposure, type, rating, and sentiment labels of bond warnings.

[0047] To maximize the potential information in the data, this feature engineering process employed statistical indicators such as logarithm, mean, mode, and extreme value. The statistical indicators showed significant predictive power based on the final model input.

[0048] Finally, an evaluation was conducted. This modeling used a semi-supervised scoring card model, which solved the problem of severely biased sample modeling and achieved good model discrimination.

[0049] The analysis process includes:

[0050] 1. Definition of target variable

[0051] The goal of this modeling is to predict the default risk of the bond issuer, so whether the issuer defaults is used as the target variable.

[0052] The default data accumulated from April 2014 to December 2019, totaling 437 records, involving 180 entities (the same entity may correspond to multiple default records). After excluding 90 entities without pre-first default warning data and 4 entities that could not match industrial and commercial information, the remaining 86 defaulting entities were used as positive samples for modeling, accounting for 0.57% of all modeling samples (15,134 entities).

[0053] 2. Data Preparation

[0054] 2.1 Construction of indicator system

[0055] To meet the modeling needs of public opinion-driven early warning of bond issuer defaults, we established an indicator system for analyzing the credit default risk of bond issuers from four perspectives: news and public opinion, industrial and commercial information, market evaluation, and external information. This indicator system comprehensively rates bond issuers across eight sub-dimensions: basic qualification information, financial and operational information, penalty information, equity pledge information, news and public opinion information, internal and external rating information, and risk correlation information. A total of 133 factors have been implemented in this modeling, as shown in Table 1:

[0056]

[0057]

[0058] Table (1)

[0059] 2.2 Feature Construction

[0060] Integrate data tables such as industrial and commercial, news and public opinion, and ratings in relational databases, and through feature engineering, realize the processing and implementation of underlying features from three aspects: statistical analysis, business judgment, and derivative construction.

[0061] 2.3 Data Verification

[0062] Check data quality to ensure feature calculations are correct. This can be divided into two parts: accuracy testing and logic testing.

[0063] 2.3.1 Accuracy Test

[0064] Statistics on missing, duplicate, field type, and abnormal conditions of each feature provide direction for subsequent outlier processing. The specific inspection method is as follows:

[0065] Missing: the number of missing values, number of explored rows, and missing proportion for each field;

[0066] Repeat: Fields with only one value (all repeats), and check the proportion of unique values ​​in each field;

[0067] Field type: whether the field type is consistent with the design;

[0068] Outliers: 3σ principle: beyond the range of the mean plus or minus three standard deviations.

[0069] 2.3.2 Logical Test

[0070] Business-defined outliers;

[0071] Calculation logic test: Through random sampling, a certain proportion of data is randomly selected, and indicators are processed using different tools, and finally the calculation results are compared.

[0072] 2.4 Data Cleaning

[0073] Missing values ​​and outliers will affect the predictive ability of factors for the final model results. By statistically analyzing the modeling samples, we can identify data noise in the modeling samples, improve data quality, and enhance model effectiveness.

[0074] 2.4.1 Missing Value Handling

[0075] Depending on the feature type, missing values ​​can be categorized as either numeric or categorical variables. When handling missing values, depending on the feature's meaning, missing values ​​can be grouped together or filled using the mean, median, or mode. Fields with severe missing values ​​(over 80% missing) are removed. For removed features, their predictive power is analyzed to determine whether they should be converted into rules for inclusion in the model. The specific handling methods are as follows:

[0076] Numerical variables: Based on the meaning of the feature and the principle of minimizing data noise, determine the missing value assignment strategy (mean, median, mode, etc.) for that feature. For example, during the modeling process, features such as the last industrial and commercial change, the last annual report disclosure, the last financial report disclosure, the establishment date, and the registration date are assigned a missing value of 99 based on their meaning. Features such as paid-in capital, registered capital, and number of employees are assigned a missing value of the mean based on their meaning. Features such as the number of branches and the number of legal person changes are assigned a missing value of 0 based on their meaning.

[0077] Categorical variables: Missing values ​​may have corresponding business implications. To maximize the preservation of information in the data, missing values ​​for categorical variables are typically grouped together and assigned a value. For example, during modeling, missing values ​​for features such as whether the company has been profitable for two or more consecutive years, whether it has suffered losses for two or more consecutive years, whether it was profitable in the most recent fiscal year, and whether it suffered losses in the most recent fiscal year are grouped together and assigned a value of -1. Missing values ​​for features such as whether a majority shareholder is a person subject to enforcement or whether the entity is a person subject to enforcement are assigned a value of 0 (indicating that the corresponding event has not occurred).

[0078] 2.4.2 Outlier Processing

[0079] This modeling does not yet generate factors that allow for business-defined outliers. For features that cannot be identified as outliers from a business perspective, Raida's rule (3σ criterion) can be used for screening. While outlier handling stabilizes the model input and helps the model capture the overall characteristics of the data, the presence of outliers can also be a true characteristic of the data. Therefore, Raida's rule was not used in this modeling.

[0080] 2.5 Training and Test Set Division

[0081] In order to verify the accuracy and stability of the model training results and to ensure good generalization ability, this modeling divided all samples into training set and test set in a ratio of 7:3.

[0082] 3. Model construction

[0083] 3.1 Exploratory Data Analysis

[0084] Through exploratory analysis of the data, the default bond issuers defined as the target variable, that is, the risk exposure sample with Y=1, totaled 86, accounting for only 0.57% of the total bond issuer samples, which is a typical unbalanced sample modeling problem.

[0085] 3.1.1 Imbalanced Sample Processing

[0086] A semi-supervised method based on PULearning (positive and unlabeled sample learning). By learning from existing labeled samples, the method finds the data point in the unlabeled sample that is closest to the labeled sample and uses it as a new labeled sample.

[0087] Since this modeling cannot obtain more bad samples, we chose to solve the sample imbalance based on PU-Learning. The following will explain in detail the principles of imbalanced sample processing and model application.

[0088] 3.2 Semi-supervised Model

[0089] Modeling positive and unlabeled samples in semi-supervised learning (Positive Unlabeled Learning) has broad application prospects in areas such as imbalanced sample processing and potential target identification, including scenarios with fewer negative labels, such as complaints and credit risk exposure. From the perspective of Positive Unlabeled Learning, known bond defaulters are only a subset of the total number of issuers with high default risk, while other issuers still have high default risk. In other words, these other samples are not purely risk-free but rather unlabeled. Therefore, the default warning model is actually modeled using positive samples (issuers that have already defaulted) and unlabeled samples (issuers with no default risk exposure).

[0090] The core idea of ​​this modeling method is that the probability of a sample being labeled is proportional to the probability of it being a positive sample. That is, from the perspective of sorting ability alone, training with positive and unlabeled samples is equivalent to training with positive and negative samples. Therefore, the classifier can be trained with positive and unlabeled samples first, and the probability of each sample being labeled output by the classifier is converted into the weight of each sample. Then, the classifier is trained again using the sample weights to obtain the probability that the sample is a positive sample (modeling target).

[0091] This modeling approach combines the aforementioned approach with a scorecard model, leveraging the ability of positive and unlabeled sample classifiers to rank risk to expand the sample size. The 5% of samples with the highest high-risk probability are used as new positive samples to train the scorecard model. This addresses the issues of having too few high-risk labels, such as difficulty distinguishing random features from truly effective ones, and poor model stability and scalability.

[0092] The scorecard model developed this time is based on the following principles and methodology: A scorecard model based on logistic regression is selected, the distribution of each feature in the positive sample is converted into an evidence weight code, and then a score is generated by combining the evidence weight and the beta in the regression coefficient. The output data-driven scorecard can intuitively reflect the information mined from the data and the model's operational logic, and can clearly provide the bond issuer scoring process and the proportion of single factor scores. The formula for converting logistic regression parameters into scorecard scores is as follows:

[0093] P0 is the baseline score of the scorecard, and PDO is the score that doubles the specified ratio. P0 and PDO are two hyperparameters of the scorecard model, which are used to control the central tendency and dispersion of the score. In this modeling, they are 60 and -10;

[0094] β is the coefficient obtained by training logistic regression, intercept is the intercept obtained by training logistic regression, and n is the number of features in the model;

[0095] Calculation of constant B:

[0096] Calculation of constant A: A = P0 + B × ln(P0);

[0097] Fixed score of the scorecard: FixedScore = AB × intercept;

[0098] Ratings for each tier:

[0099] The steps for building a semi-supervised model are as follows:

[0100] (1) Training a classifier for whether a sample is labeled: Train the XGBoost model to learn whether a sample is labeled. Grid search is used to adjust the parameters with AUC as the target. The optimal hyperparameters are shown in Table 2:

[0101] Colsample_bytree 1 Learning_rate 0.01 Max_depth 10 N_estimators 200

[0102] Table (2)

[0103] XGBoost is an optimized distributed gradient boosting library designed to be efficient, flexible, and portable. It implements machine learning algorithms under the GradientBoosting framework. XGBoost provides parallel tree boosting (also known as GBDT, GBM) that can solve many data science problems quickly and accurately.

[0104] The AUC (Area Under Curve) is defined as the area under the ROC curve and the coordinate axes. Obviously, this area will never be greater than 1. Since the ROC curve generally lies above the line y = x, the AUC range is between 0.5 and 1. The closer the AUC is to 1.0, the more reliable the detection method is; a value of 0.5 indicates the lowest reliability and no practical value.

[0105] (2) Probability calibration: CalibratedClassifierCV is used to perform probability calibration, calibrating the output of XGBoost to the probability of the approximate standard. The hyperparameters used are method_calibrated = isotonic, cv = 3, as shown in Table 3:

[0106] Method_Calibrated Isotonic CV 3

[0107] Table (3)

[0108] (3) Construct an expanded modeling target: Use the top 5% samples ranked by calibrated probability and take the union with the original negative labels as the modeling target for the subsequent training scorecard;

[0109] (4) Imbalanced sample weighting: This ensures that the model does not underestimate the misclassification of positive samples due to the fact that the positive samples only account for 5%. The weights are calculated using sklearn.utils.class_weight.

[0110] (5) Feature discretization: Use chi-square binning to convert all features into ordinal categorical variables. The hyperparameters of chi-square binning are shown in Table 4:

[0111] Max_intervals 10 Min_intervals 5 Initial_intervals 100

[0112] Table (4)

[0113] (6) Predictive power and collinearity analysis: Analyze the degree of correlation between features and modeling targets, as well as the collinearity between features, to screen high-quality features that can be included in the model;

[0114] (7) Manually optimize feature interpretability: Check each high-quality feature that can be included in the model, analyze whether the frequency distribution and response rate distribution of each value can be explained in the business, and whether it is likely to be caused by random fluctuations in the data. Adjust the grouping of features accordingly, and check whether the validation set has the same trend as the training set. Features that cannot be explained, are likely to be caused by random fluctuations in the data, or have inconsistent trends between the training set and the validation set cannot be included in the model;

[0115] (8) Evidence weight conversion and score training: The features are encoded with evidence weights and then the scorecard model is trained. The hyperparameters are shown in Table 5:

[0116] Max_intervals 10 Min_intervals 5 Initial_intervals 100

[0117] Table (5)

[0118] (9) Scorecard adjustment: manually review the scoring rules and correct a few rules that do not match the response rate distribution results;

[0119] The English involved in the semi-supervised model modeling steps refers to the parameter settings in the code.

[0120] The final semi-supervised default model input indicators are as follows Table 6:

[0121]

[0122]

[0123] Table (6)

[0124] 4. Model Evaluation

[0125] 4.1 Evaluation Method

[0126] The scenario of early warning for the default of bond-issuing entities is not a conventional classification problem with positive and negative samples, but a semi-supervised problem with labeled positive samples + unlabeled samples. It is difficult for the sample ratio to meet the training requirements of traditional classification models. In the samples for this modeling, the number of defaulting entities is too small, and the remaining entities are actually a mixture of entities with high risks but not yet defaulted and entities with lower risks. The precision metric will regard entities that are predicted by the model as high-risk but not yet defaulted as prediction errors. However, in fact, such entities are truly approaching defaulting entities in terms of risk characteristics, only the risks have not been exposed or there are other factors outside the model that have caused them not to default yet. Therefore, the traditional precision metric is no longer applicable to the scenario of early warning for the default of bond-issuing entities. In this modeling, the AUC is used to evaluate the overall risk ranking ability of the model, and the KS is used to evaluate the ability of the model to distinguish between positive and negative samples.

[0127] AUC (Area Under the ROC Curve): It tests the ranking ability of the model. It is recommended that the AUC value be above 0.7. The higher the AUC, the better the classification effect of the model, and the greater the probability that default samples are ranked in front of non-default samples.

[0128] 0.5 < AUC < 1, better than random guessing, with predictive value; AUC = 0.5, the same as random guessing, without predictive value;

[0129] K-S statistic: It tests the discrimination ability of the model. It is recommended that the KS value be above 0.40.

[0130] 0.4 < KS, the model has good discrimination ability; 0.2 < KS ≤ 0.4, the model has average discrimination ability; KS ≤ 0.2, the model has poor discrimination ability.

[0131] 4.2 Semi-supervised Model Evaluation

[0132] As shown in Tables (7) and (8), finally, the AUC on the test set is 0.9617, close to 1; the KS is 0.7779, greater than 0.4, indicating that the semi-supervised model has good risk ranking ability and discrimination ability for predicting the default probability of bond-issuing entities. At the same time, after sorting in descending order of the warning scores, at the threshold of the top 2%, the recall rate reaches 88.37%, which also reflects that the model has good predictive ability for default risks.

[0133] Evaluation indicators Full sample AUC 0.9611 KS 0.8191

[0134] Table (7)

[0135] Abnormal level in descending order Cumulative recall rate Top 1% 74.42％ Top 2% 81.40％ Top 5% 87.21％ Top 10% 90.70％ all 100.00％

[0136] Table (8)

[0137] The present invention provides a method for predicting default of a bond issuer based on a semi-supervised model, comprising:

[0138] Whether the bond issuer defaults is used as the target variable;

[0139] Obtain the main data of the bond issuer's news and public opinion information, industrial and commercial information, market evaluation information, and external information. Based on this main data, an indicator system for the bond issuer's credit default risk is constructed. The indicator system comprehensively rates the bond issuer from eight sub-dimensions, including basic qualification information, financial and operating information, penalty information, equity pledge information, news and public opinion information, internal and external rating information, and risk association information.

[0140] Through feature engineering, we integrate data tables for business, news, and ratings in relational databases. We process underlying features from three perspectives: statistical analysis, business judgment, and derivative construction. We generate underlying factors and test data quality through accuracy and logic testing to ensure the correctness of feature calculations, filter out outliers, remove missing values, improve data quality, and enhance predictive capabilities. To uncover the potential information in the bond issuer's data, feature engineering uses statistical indicators such as logarithm, mean, mode, and extreme value.

[0141] A semi-supervised model is established based on a combination of the unlabeled sample weighting method and the scorecard model. The risk ranking ability of the positive sample and unlabeled sample classifier is used to expand the scale of the positive sample. The samples with the highest high-risk probability are used as new positive samples to train the scorecard model. This solves the problems of too few high-risk labels, such as difficulty distinguishing between random and truly effective features, and weak model stability and scalability. The scorecard model is based on logistic regression. It converts the distribution of each feature in the positive sample into an evidence weight code, and then generates a score by combining the evidence weight and the β in the regression coefficient. The output data-driven scorecard model reflects the information mined from the data and the model's operational logic, clearly presenting the bond issuer's scoring process and the proportion of single factor scores.

[0142] Judge and predict the default risk of the bond issuer based on the semi-supervised model.

[0143] By utilizing the positive sample and unlabeled sample learning methods in semi-supervised learning, the scale of positive samples is expanded and the originally severely biased modeling samples are corrected. On the one hand, this addresses the possibility of labeled samples existing in unlabeled samples. On the other hand, it enables the model to better learn the characteristics of bad samples and reduces the risk of the model fitting more noise due to sample imbalance.

[0144] The method for predicting the default of a bond issuer based on a semi-supervised model provided by the present invention includes the steps of establishing a semi-supervised model:

[0145] Through grid search, we adjust the parameters with AUC as the target and train the XGBoost model to obtain a classifier that can identify whether the sample is labeled.

[0146] Use the calibration classifier to perform probability calibration, and use the output of XGBoost calibration as the probability of the approximate standard;

[0147] Use the calibrated samples and the original negative labels as the modeling target for the subsequent training scorecard;

[0148] Calculate weights using balanced sample weighting;

[0149] Use chi-square binning to convert all features into ordinal categorical variables;

[0150] Analyze the correlation between features and modeling targets, as well as the collinearity between features, to select high-quality features that can be included in the model;

[0151] Manually optimize feature interpretability. Check each high-quality feature that can be included in the model. Analyze whether the frequency distribution and response rate distribution of each value can be explained from a business perspective and whether they may be due to random fluctuations in the data. Adjust the feature grouping accordingly and check whether the validation set has the same trend as the training set. Features that cannot be explained, are likely due to random fluctuations in the data, or have inconsistent trends between the training and validation sets cannot be included in the model.

[0152] The features are encoded by weight of evidence and then the scorecard model is trained;

[0153] The scoring rules were manually reviewed and a few rules that did not match the response rate distribution results were corrected.

[0154] The method provided by the present invention for predicting the default of a bond issuer based on a semi-supervised model includes testing the discrimination ability of the semi-supervised model through a KS evaluation model. When KS>0.4, the discrimination ability is good.

[0155] The method provided by the present invention for predicting the default of a bond issuer based on a semi-supervised model includes testing the ranking ability of the model through AUC. The range of AUC is AUC>0.7. The higher the AUC, the better the model classification effect, and the greater the probability that the default sample is ranked before the non-default sample.

[0156] The present invention also provides a device for predicting the default of a bond issuer based on a semi-supervised model. The device for predicting the default of a bond issuer based on a semi-supervised model includes:

[0157] Memory, processor, communication bus and a semi-supervised model stored in the memory to predict the default procedure of the bond issuer,

[0158] The communication bus is used to realize the communication connection between the processor and the memory;

[0159] The processor is used to execute a procedure for predicting the default of a bond issuer based on a semi-supervised model, so as to implement any of the steps of the method for predicting the default of a bond issuer based on a semi-supervised model as described above.

[0160] The present invention also provides a computer-readable storage medium storing executable instructions, and the storage medium storing a program for predicting the default of a bond issuer based on a semi-supervised model. When the program for predicting the default of a bond issuer based on a semi-supervised model is executed by a processor, the steps of any one of the above-mentioned methods for predicting the default of a bond issuer based on semi-supervised machine learning are implemented.

[0161] The above descriptions are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention's description and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A method for predicting default of a bond issuer based on a semi-supervised model, characterized by: include: S1: Obtain the principal data of the bond issuer, including news and public opinion information, industrial and commercial information, market evaluation information, and external information, and construct an indicator system for the credit default risk of the bond issuer based on the principal data; S2: Through feature engineering, we process and implement the underlying features from three aspects: statistical analysis, business judgment, and derivative construction, and generate the underlying factors. S3: Establish a semi-supervised model based on the combination of unlabeled sample weighting method and scorecard model; S4: Predicting the default risk of bond issuers based on a semi-supervised model; The establishment of the semi-supervised model of S3 includes the following steps: S21: Optimize the parameters with AUC as the target through grid search and train the XGBoost model to obtain a classifier that can identify whether the sample is labeled. S22: Use the calibration classifier to perform probability calibration, calibrating the output of XGBoost to approximately standard probabilities. The hyperparameters used are method_calibrated=isotonic, cv=3; S23: Use the calibrated probability to sort the top 5% of samples and take their union with the original negative labels as the modeling target for the subsequent training scorecard. The modeling target is the probability of positive samples, which is used to solve the problem of too few high-risk labels making it difficult to distinguish between random and truly effective features. S24: Calculate weights using balanced sample weighting; S25: Use chi-square binning to convert all features into ordinal categorical variables; S26: Analyze the correlation between features and modeling targets, as well as the collinearity between features, to select high-quality features that can be included in the model; S27: Manually optimize feature interpretability. Check each high-quality feature that can be included in the model. Analyze whether the frequency distribution and response rate distribution of each value can be explained from a business perspective and whether it may be due to random fluctuations in the data. Adjust the feature grouping accordingly and check whether the validation set has the same trend as the training set. S28: Encode the features by weight of evidence and then train the scorecard model; S29: Manually review the scoring rules and correct a few rules that do not match the response rate distribution results; The AUC of S21 is in the range of AUC>0.

7.

2. The method for predicting bond issuer default based on a semi-supervised model according to claim 1, characterized in that: The S1 index system rates the bond issuer based on basic qualification information, financial and operational information, penalty information, equity pledge information, news and public opinion information, internal and external rating information, and risk-related information.

3. The method for predicting bond issuer default based on a semi-supervised model according to claim 1, characterized in that: The S2 uses statistical indicators such as logarithm, mean, mode and extreme value to mine the potential information of the bond issuer data.

4. The method for predicting bond issuer default based on a semi-supervised model according to claim 1, characterized in that: The S3 scoring card model is a scoring card model based on logistic regression. It converts the distribution of each feature in the positive sample into evidence weight coding, and then generates a score by combining the evidence weight and β in the regression coefficient. The output data-driven scoring card model reflects the information mined from the data and the model's operational logic, and gives the scoring process of the bond issuer and the proportion of single factor scores.

5. The method for predicting bond issuer default based on a semi-supervised model according to claim 1, characterized in that: The semi-supervised model of S3 was tested for discrimination ability using the KS evaluation model, with KS>0.

4.

6. A device for predicting default of a bond issuer based on a semi-supervised model, characterized by: The device for predicting bond issuer default based on the semi-supervised model includes: Memory, processor, communication bus and a semi-supervised model stored in the memory to predict the default procedure of the bond issuer, The communication bus is used to realize the communication connection between the processor and the memory; The processor is used to execute the program for predicting the default of the bond issuer based on the semi-supervised model to implement the steps of the method for predicting the default of the bond issuer based on the semi-supervised model as described in any one of claims 1 to 5.

7. A computer-readable storage medium storing executable instructions, characterized in that: The storage medium stores a program for predicting the default of a bond issuer based on a semi-supervised model. When the program for predicting the default of a bond issuer based on a semi-supervised model is executed by a processor, the steps of the method for predicting the default of a bond issuer based on a semi-supervised model described in any one of claims 1 to 5 are implemented.

Images