Early warning method in credit service of automobile dealer and storage medium

Abstract

The invention discloses an early warning method in credit service of an automobile dealer. According to the early warning method, a machine learning model early warning module and an automatic empirical rule early warning module are adopted to carry out monitoring and early warning on monitoring data and output a corresponding early warning result; wherein the machine learning model early warning module adopts a machine learning model to analyze dealer monitoring data and outputs a corresponding early warning result; and the automatic empirical rule early warning module analyzes the monitored dealer monitoring data according to a preset rule and outputs a corresponding early warning result. The multi-dimensional data of the dealer is used as a decision basis, and at least two early warning methods are adopted to carry out risk early warning, so that the monitoring efficiency and quality in the credit service of the automobile dealer are improved, and accurate early warning is realized.

Description

Technical Field

[0001] The present invention relates to the field of early warning of automobile credit risks, and particularly to a method and a storage medium for early warning during the credit granting of automobile dealers. Background Art

[0002] In recent years, with the rapid development of artificial intelligence technologies such as machine learning and deep learning, more and more enterprises and institutions have begun to explore the application of artificial intelligence and big data technologies in actual business in order to reduce costs and improve efficiency. In the field of automotive finance, corporate credit for dealers is an important business. After a dealer is granted credit, it is necessary to comprehensively monitor its internal and external risks.

[0003] The traditional methods for monitoring dealers during the loan period in the automotive finance industry mainly rely on manual investigation and passive reception, and perform poorly in terms of risk identification effect, manpower input, and efficiency. With the continuous increase in the number of corporate credit network applications in automotive finance, the volume of corporate credit business has also shown a relatively fast growth trend. Therefore, higher requirements are put forward for the risk identification ability and timeliness of corporate credit monitoring during the loan period. It is urgent to build an automated early warning system during the loan period. On the one hand, it can identify relevant risks of customers during the loan period in advance, and on the other hand, it can capture and warn risk customers more accurately. Summary of the Invention

[0004] The purpose of the present invention is to overcome the deficiencies of the prior art, and to provide a method and a storage medium for early warning during the credit granting of automobile dealers. By using multi-dimensional data of dealers as the decision-making basis, a rule model system for automated monitoring is constructed to improve the efficiency and quality of monitoring during the credit granting of automobile dealers and achieve accurate early warning.

[0005] To achieve the above purpose, the technical solution adopted by the present invention is: a method for early warning during the credit granting of automobile dealers. The early warning method uses a machine learning model early warning module and an automated experience rule early warning module to monitor and warn the monitoring data and output corresponding early warning results. Among them, the machine learning model early warning module uses a machine learning model to analyze the dealer monitoring data and output corresponding early warning results. The automated experience rule early warning module analyzes the monitored dealer monitoring data according to preset rules and outputs corresponding early warning results.

[0006] In the machine learning model early warning module, the internal and external data indicators of the granted dealers are used as the model input indicators, including internal overdue information, default information, and inventory information.

[0007] The automated empirical rule warning module includes a classification rule unit and a monitoring rule unit. The classification rule unit automatically classifies dealers with credit contracts according to the set clear risk conditions. The monitoring rule generates warning items and investigation tasks for dealers with credit contracts according to the set non-clear risk conditions, and sends the warning items and investigation tasks to manual investigation to complete the risk classification of dealers.

[0008] The risk situation of a single data dimension is warned by the automated empirical rule warning module; the risk situation of multi-dimensional data is warned by the machine learning model warning module.

[0009] The data dimensions input into the machine learning model warning module and the automated empirical rule warning module include one or a combination of industry and commerce, judiciary, finance, external credit investigation, vehicle-mounted TBOX information, internal black and gray lists, internal repayment, internal compliance information, vehicle pick-up, sales and inventory information.

[0010] The machine learning model in the machine learning model warning module adopts the LR model. The multi-dimensional dealer data is cleaned, crossed, and feature engineered to generate dealer dimension indicators and sent into the machine learning model, and the LR model outputs a warning score.

[0011] According to the warning score output by the LR model, a risk warning is directly issued to dealers with a score lower than the set standard; a risk warning is issued to dealers with a score higher than the set standard and with default extension record information.

[0012] A standard threshold is preset in the automated empirical rule warning module, and the index parameters input into the automated empirical rule warning model are compared with the corresponding standard threshold, and the output risk level is judged based on the comparison result.

[0013] A storage medium stores at least one instruction, and the instruction is loaded and executed by a processor to implement the in-loan warning method.

[0014] The advantages of the present invention are as follows: By integrating machine learning algorithms, mathematical statistics methods and business expert experience, a rule system of in-loan warning models is constructed. For in-loan dealer risk matters, on the one hand, potential risks are predicted in advance through the warning model, reminding business personnel to take corresponding risk prevention measures in advance to avoid risk losses as much as possible; on the other hand, risk matters are processed in time through warning rules to reduce risk losses as much as possible. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The following briefly describes the content expressed in each drawing of the specification of the present invention and the marks in the drawings:

[0016] Figure 1This is the control flow chart of the in - loan early warning method for automobile dealers in the present invention. Detailed implementation manners

[0017] The following is a further detailed description of the specific implementation manners of the present invention by describing the optimal embodiments with reference to the accompanying drawings.

[0018] This solution is mainly aimed at the in - loan risk early warning of automobile dealers by financial companies. Reliable risk prediction is achieved through two risk assessment methods in this solution, which is beneficial to timely discover risks, give early warnings and handle them in a timely manner, and provide lease early warning reminders for automobile financial companies. The specific solution is as follows:

[0019] As Figure 1 shown, an in - loan early warning method for automobile dealers provided in this embodiment uses a machine learning model early warning module and an automated experience rule early warning module to monitor and give early warnings to monitoring data and output corresponding early warning results. Among them, the machine learning model early warning module uses a machine learning model to analyze the dealer monitoring data and output corresponding early warning results; the automated experience rule early warning module analyzes the monitored dealer monitoring data according to preset rules and outputs corresponding early warning results.

[0020] Based on the monitoring index data of automobile dealers, early warning analysis is respectively carried out through the machine learning model and the automated experience rule model to obtain the analyzed early warning results, which can more accurately and reliably achieve early warning. Both the machine learning model early warning module and the automated experience rule early warning module will input the monitoring index data of the monitored automobile dealers, and based on these monitoring index data, risk early warning is realized. The monitoring index data can be set and adjusted according to different risk early warning requirements.

[0021] In the machine learning model early warning module, it uses the internal and external data indexes of the credited dealers as the model input indexes, including internal overdue information, default information, purchase - sale - inventory information, etc. These data index information are all data indexes representing the risks of dealers. These indexes are collected and sent into the machine learning early warning model to output the risk early warning results.

[0022] The automated experience rule early warning module includes a classification rule unit and a monitoring rule unit. Among them, the classification rule unit automatically classifies the risks of dealers with credit contracts according to the set clear risk conditions; the monitoring rule generates early warning items and investigation tasks for dealers with credit contracts according to the set non - clear risk conditions, and sends the early warning items and investigation tasks to manual investigation to complete the risk classification of dealers. The preset clear risk conditions include conditions such as the financial status of the enterprise and the overdue situation of the enterprise. According to these conditions, the dealers are monitored to achieve the purpose of automatically outputting early warning results.

[0023] Among them, the risk situation of a single data dimension is warned by the automated experience rule warning module; the risk situation of comprehensive multi-dimensional data is warned by the machine learning model warning module. That is, a variety of data of automobile dealers are collected, each kind of data corresponds to a dimension, and the automated experience rule warning module judges the risk for each kind of data and gives a warning result; while the machine learning model comprehensively analyzes and judges a variety of data to give a warning result.

[0024] The multi-dimensional data input into the machine learning model warning module and the automated experience rule warning module includes, but is not limited to, industrial and commercial information, judicial information, financial information, external credit investigation information, vehicle-mounted TBOX information, internal black and gray list information, internal repayment information, internal compliance information, vehicle pick-up, sales and inventory information. Through these information, the purpose of risk warning is achieved by means of models or rules.

[0025] In this embodiment, the machine learning model in the machine learning model warning module adopts the LR model. After building the LR model, it is necessary to train and optimize the LR model to obtain a trained LR model that can output risk results. The multi-dimensional dealer data is cleaned, crossed, and feature engineered to generate dealer dimension indicators and sent into the machine learning model, and the LR model outputs a warning score. Then, according to the warning score output by the LR model, a risk warning is directly issued to the dealer whose score is lower than the set standard; a risk warning is issued to the dealer whose score is higher than the set standard and has default extension record information. The standard score threshold can be set, and then the score output by the LR model of each dealer is compared with the standard score threshold. When it is less than the standard score threshold, a risk warning is directly output to achieve the threshold reminder for low scores; when the score output by the LR is higher than the standard value but it is detected that the dealer has default extension record information, it means that the dealer has risks at this time, and a corresponding reminder is also issued at this time.

[0026] In the automated experience rule warning module, standard thresholds are preset, and the index parameters input into the automated experience rule warning model are compared with the corresponding standard thresholds, and the output risk level is judged based on the comparison result. For example, multiple risk levels are set, and the risk level is judged according to the comparison with the threshold, and the risk level is used as the warning result to facilitate the warning of the enterprise after granting credit to the dealer.

[0027] This embodiment also provides a storage medium, in which at least one instruction is stored, and the instruction is loaded and executed by a processor to implement the in-loan warning method in the above embodiment.

[0028] In this solution, the warning rule model warning method for in-loan monitoring of automobile dealer credit granting includes,

[0029] S1: The loan early warning model system is divided into: machine learning model early warning and automated experience rule early warning;

[0030] S2: Introduce comprehensive and authentic internal and external data dimensions, especially internal overdue, default and inventory-related indicators, to improve the scientific rationality of model / rule warning results;

[0031] S3: Combine business and risk control needs with big data analysis technology to mine indicators for early warning models and apply machine learning algorithms to establish early warning models;

[0032] S4: Build a model warning strategy based on the warning model results and other business rule requirements to improve the overall effect of the warning model;

[0033] S5: Automated experience rules include classification rules and monitoring rules, which can conduct comprehensive monitoring of loan customers in a timely manner;

[0034] S6: Classification rules are to automatically classify the risks of dealers with credit contracts according to the set business conditions;

[0035] S7: Monitoring rules generate early warning items and troubleshooting tasks for dealers with credit contracts according to the set business conditions, and classify dealers by risk after manual troubleshooting;

[0036] The entire dealer loan early warning model consists of two parts: machine learning model early warning and automated experience rule early warning. The machine learning model early warning consists of early warning scoring model and model early warning strategy. The automated experience rule early warning consists of classification rules and monitoring rules.

[0037] In step S1, the risk situation of a single data dimension is warned through automated empirical rules, and the risk situation of comprehensive multi-dimensional data is warned through a machine learning model.

[0038] In step S2, its comprehensive, authentic and reliable data dimensions include not only general industrial and commercial, judicial, financial, external credit, vehicle TBOX and other information, but also internal black and gray lists, internal repayments, internal compliance, vehicle pick-up and sales inventory and other authentic and reliable business information, laying a high-quality data foundation for the construction of rules / models;

[0039] In step S3, the multi-dimensional basic data is cleaned, crossed, and feature engineered to derive dealer dimension indicators, build a machine learning model, output a warning score, and evaluate dealer risks from a business and risk perspective.

[0040] In step S4, the dealers are grouped according to the risk situations of dealers in each score segment. Dealers with the lowest specific proportion of model scores are directly warned, and dealers with relatively low model scores and records of default extensions are supplemented with warnings, thus overall improving the effect of the warning model.

[0041] In step S5, the rules for severe risks are defined as classification rules. If a rule is hit, a medium / high risk label is directly and automatically assigned; the rules for general risks are defined as monitoring rules. If a rule is hit, a low / medium / high risk label is manually assigned after manual investigation.

[0042] In step S6, the warning conditions for the classification rules include default, overdue, and model warning times conditions exceeding specific thresholds. The thresholds are obtained through data analysis. After being hit, high / medium two risk levels are automatically assigned.

[0043] In step S7, the monitoring rules are based on information such as industry and commerce, judiciary, credit investigation, business risks, and internal defaults. After being hit, the risk level is to be determined. After manual investigation, it may be assigned one of the four risk levels: high / medium / low / none.

[0044] The early warning rule model system for the in - credit monitoring of auto dealers' credit includes machine learning model early warning and automated experience rule early warning. Machine learning model early warning focuses on predicting risks in advance, and automated experience rule early warning focuses on timely reminding of post - event risks. The two complement each other. Machine learning model early warning conducts risk assessment on all credit - granted dealers from perspectives such as the enterprise's own credit investigation, the actual controller's credit investigation, and business information, and gives early warning reminders to dealers with relatively low scores and the core reasons, facilitating business personnel to conduct risk investigations in a timely manner and discover and handle risk anomalies in advance. Automated experience rule early warning includes classification rules and monitoring rules. Classification rules classify the risks of credit - granted dealers from aspects such as default, overdue, and historical warning situations, and automatically mark high / medium 2 risk levels. Monitoring rules give early warnings to credit - granted dealers from perspectives such as industry and commerce, judiciary, credit investigation, and internal risks, send early warning information and details to business personnel, and business personnel conduct risk investigations and confirmations and mark potential risks for disposal.

[0045] The early warning model system for the in - credit warning of auto dealers' credit includes two major modules: machine learning model early warning and automated experience rule early warning.

[0046] Machine learning model early warning: It is mainly divided into three stages. The first stage is model construction, and the second stage is strategy development.

[0047] The first stage (model construction):

[0048] Screen and construct effective samples for the early warning scoring model according to business requirements. By comprehensively considering multiple dimensions such as the seasonal changes in dealers' operations, the time range of available data dimensions, reasonable performance periods and observation periods, sample size, and the definition of good and bad samples, select the samples and related basic data sets for modeling.

[0049] Construct derivative variables from multiple dimensions. The derivative dimensions include but are not limited to the short, medium, and long-term profitability, repayment willingness, debt-servicing ability, and cooperation creditworthiness of the dealer's operation process, and the personal credit assessment of the actual controller.

[0050] Screen the variables. Eliminate variables from the perspectives of IV value, univariate significance, multivariate significance, multicollinearity, correlation, and stability, and reselect some of the eliminated variables according to expert experience, so as to select alternative variables as the shortlist for modeling. The variable screening process and the steps and criteria adopted are as follows: Screen out variables with incorrect index processing logic and certificate-related variables according to the original data situation; (1) Conduct a preliminary screening of the variables and view the distribution of each variable in the long list; (2) Delete variables with a missing rate greater than 90%, a concentration greater than 0.95, and an IV value less than 0.02, that is, screen out variables with a relatively complete variable value distribution and the ability to distinguish; (3) Conduct a rough binning on the basis of the preliminarily screened variables and observe the bad sample rate and trend of the variable binning; (4) Delete some variables with high correlation, and select several variables with greater discrimination ability (referring to the IV value) from different transformations of the same type of variables as alternative variables; (5) Delete variables with a PSI greater than 0.1 according to the stability assessment in the cross-time validation set; (6) Based on the above-screened variables, bin the remaining variables again and preliminarily adjust the variable binning, and then adjust the bad customer ratio and monotonicity between different variable bins; (7) Compare the modeling effects, iterate the modeling variables multiple times, and select variables optimally. At the same time, according to expert experience, reselect some of the deleted variables and delete some variables that do not conform to the business interpretation, and finally calculate the VIF for verification.

[0051] Build a model based on the variables. Encode the variables. For variables with business interpretability, select the woe transformation method for encoding. Bin the variables in the optimal way of information entropy gain, so that the bad sample rate trend shows monotonic interpretability, the proportion of each bin is relatively uniform, and the IV value is optimal. Calculate the woe and IV value as follows:

[0052]

[0053]

[0054] where i is the number of bins, y i and y T are the number of bad samples and the total number of bad samples in this bin respectively, n iWith n T They are the number of samples and the total number of samples in this bin, respectively.

[0055] Subsequently, the variables are transcoded according to the woe of their respective bins, and the information of the transcoded variables is input into the LR model for training. The LR model formula is as follows:

[0056]

[0057] Where x is the woe value of the indicator variable, and w T is the variable coefficient.

[0058] Again, calculate the scoring results by calculating the parameters of the model training and the woe values of the indicators as follows:

[0059]

[0060] Where A and B are constants, θ is the training coefficient of the model corresponding to each indicator, ω is the woe value of the bin corresponding to each indicator, δ is a 0, 1 logical variable, taking 0 means not taking this bin, n represents the nth indicator, and j represents the jth bin.

[0061] Finally, process each indicator in the data warehouse, obtain data at the decision engine development interface, and deploy the model at the decision engine.

[0062] The second stage (policy development):

[0063] The early warning scoring model provides a refined credit risk measurement tool and improves the short-term credit risk prediction ability. Through the analysis of the scoring model distribution and comprehensive business evaluation, the post-loan customer group can be quantitatively risk-segmented through the early warning score, so as to accurately focus on dealers with high credit risks, achieve the goal of warning a small number of dealers and capturing most risks, and improve the differentiated post-loan management mechanism.

[0064] After sorting the modeling sample scores in ascending order, perform equal-frequency binning on the full-scale samples, sequentially select the bin cut-off points, calculate the sample capture rate and the cumulative bad sample capture rate under this cut-off point, divide this value by the bad sample rate of the overall sample, and calculate the lift value. Communicate with the business regarding multiple groups of sample cut-off rates, lift values, and the business, select the sample capture rate that the business can accept and a reasonable lift value, and the score value corresponding to this sample cut is the threshold set for the model early warning strategy. If the splitting ability of a single score for the sample is not significant enough, the model score can be crossed with other variables not included in the model to specify a combined strategy, and the strategy formulation method is the same as above. Finally, process each indicator in the data warehouse, obtain data at the decision engine development interface, and deploy the strategy at the decision engine.

[0065] Automated experience rule warning: It is mainly divided into two stages. The first stage is the development and deployment of classification rules, and the second stage is the development and deployment of monitoring rules.

[0066] The first stage (development and deployment of classification rules):

[0067] The dealers hit by the classification rules have clear risks. After being hit, they are automatically marked as high / medium risks without manual investigation. According to expert experience, the direction of clear design risks is given, including but not limited to: multiple overdue payments, multiple extensions, multiple defaults, multiple historical warnings, etc. Develop variables around this direction, design rules for different thresholds of variables, define clear risk bad samples, and calculate the rule hit rate and lift value (bad sample rate of hit rules / bad sample rate of all samples). (Finally, communicate with the downstream business system that displays warning information about the output format, deploy the rules in the decision engine, and develop the front end and back end in the downstream business system.

[0068] The second stage (development and deployment of monitoring rules):

[0069] The monitoring rules give warnings from internal risk perspectives such as enterprise industrial and commercial information, enterprise judicial information, enterprise credit investigation, personal judicial information of enterprise directors, supervisors and senior managers, personal credit investigation of enterprise directors, supervisors and senior managers, and mild overdue payments / extensions / defaults. Triggering the corresponding rules means giving a warning, and business personnel conduct risk investigation and confirmation based on the warning information to solve subsequent problems in a timely manner.

[0070] First, test and introduce external data, select the required dimensions from multiple dimensions of external data, and develop data docking interfaces and data processing scripts. Secondly, evaluate and screen monitoring rules from the perspectives of data and implementation to ensure the feasibility of monitoring rules. Then, backtest and analyze and verify the monitoring rules from multiple perspectives such as warning volume, warning rate, ability to capture bad samples, and correlation and concurrency of multiple rules, explore the reasonable thresholds of single rules and the combined effects of multiple rules, and improve the effectiveness of monitoring rules for risk warning. Finally, communicate with the downstream business system that displays warning information about the output format, deploy the rules in the decision engine, and develop the front end and back end in the downstream business system.

[0071] Obviously, the specific implementation of the present invention is not limited by the above methods. As long as various non-substantive improvements are made by adopting the method concept and technical solution of the present invention, they are all within the protection scope of the present invention.

Claims

1. A car dealer credit early warning method, characterized by: The early warning method uses a machine learning model early warning module and an automated empirical rule early warning module to monitor and warn the monitoring data and output corresponding early warning results; The machine learning model early warning module uses a machine learning model to analyze dealer monitoring data and output corresponding early warning results; The automated empirical rule warning module analyzes the monitored dealer monitoring data according to preset rules and outputs corresponding warning results.

2. The automobile dealer credit early warning method as claimed in claim 1, characterized in that: In the machine learning model early warning module, internal and external data indicators of the authorized dealers are used as model input indicators, including internal overdue information, default information, and purchase, sales and inventory information.

3. The automobile dealer credit early warning method as claimed in claim 1, characterized in that: The automated empirical rule warning module includes a classification rule unit and a monitoring rule unit, wherein the classification rule unit automatically classifies the risks of dealers with credit contracts according to the set explicit risk conditions; the monitoring rule generates warning items and troubleshooting tasks for dealers with credit contracts according to the set implicit risk conditions and sends the warning items and troubleshooting tasks to manual troubleshooting to complete the risk classification of dealers.

4. A car dealer credit early warning method according to any one of claims 1 to 3, characterized in that: Use the automated experience rule warning module to warn of risks in a single data dimension; The risk situation of comprehensive multi-dimensional data is warned through the machine learning model warning module.

5. A car dealer credit early warning method according to any one of claims 1 to 3, characterized in that: The data dimensions input into the machine learning model warning module and the automated experience rule warning module include one or more combinations of industrial and commercial, judicial, financial, external credit, vehicle TBOX information, internal black and gray lists, internal repayments, internal compliance information, and vehicle pick-up, sales, and inventory information.

6. A car dealer credit early warning method according to any one of claims 1 to 3, characterized in that: The machine learning model in the machine learning model early warning module adopts the LR model. The multi-dimensional dealer data is cleaned, crossed, and feature-engineered to generate dealer dimension indicators which are sent into the machine learning model, and the LR model outputs the early warning score.

7. The method for early warning of credit approval for automobile dealers as claimed in claim 6, characterized in that: Based on the warning score output by the LR model, a risk warning is directly issued to dealers whose scores are lower than the set standard; a risk warning is issued to dealers whose scores are higher than the set standard and have a record of default extension information.

8. A car dealer credit early warning method according to any one of claims 1 to 3, characterized in that: The standard threshold is pre-set in the automated experience rule warning module, the indicator parameters input into the automated experience rule warning model are compared with the corresponding standard threshold, and the output risk level is determined based on the comparison result.

9. A storage medium, characterized in that: The storage medium stores at least one instruction, and the instruction is loaded and executed by the processor to implement the method according to any one of claims 1-8.

Images