CN112508689A

CN112508689A - Method for realizing decision evaluation based on multiple dimensions

Info

Publication number: CN112508689A
Application number: CN202110135844.7A
Authority: CN
Inventors: 江海; 李开宇
Original assignee: Sichuan XW Bank Co Ltd
Current assignee: Sichuan XW Bank Co Ltd
Priority date: 2021-02-01
Filing date: 2021-02-01
Publication date: 2021-03-16

Abstract

The invention discloses a method for realizing decision evaluation based on multiple dimensions, which belongs to the technical field of intelligent decision mining and aims at solving the problems that in the prior art, the credit earnings are quantized only from a risk perspective, the refined operation of a client is difficult to realize, the credit earnings are quantized only from the risk perspective, and the pricing strategy is difficult to adjust from a business perspective; packaging into API interface service, and collecting data required by the credit authorization of the client when the client applies for the credit authorization in the front-end interactive tool; then obtaining output values corresponding to different models, obtaining personalized pricing of the customer according to the output values, and calculating expected income of the customer; and displaying the personalized pricing through a front-end interactive tool and feeding back the personalized pricing to a client. The invention is used for decision evaluation.

Description

Method for realizing decision evaluation based on multiple dimensions

Technical Field

The invention belongs to the technical field of intelligent decision mining, and particularly relates to a method for realizing decision evaluation based on multiple dimensions.

Background

Under the era of rapid development of big data and artificial intelligence technologies, the big data and the intelligent technologies are applied to the financial field, the credit products can be quantized from multiple dimensions, and the credit products are finely priced by combining intelligent quantization indexes with the service direction, the enterprise development stage, the market scale, the passenger group positioning and the like.

As shown in fig. 1, the pricing technical solution of credit products in the prior art is as follows:

the transverse process is a model construction process:

the method comprises the following steps: data acquisition: collecting information data of a client;

step two: data processing and integration: summarizing and sorting all the collected data;

step three: training a wind control model: performing model training and parameter tuning on the risk data to finally obtain a trained model file;

step four: model service: packaging the obtained model file into API interface service;

the longitudinal process is a user credit application pricing process:

step five: the user applies for credit: in the process of applying for credit by a client, collecting data information of the client;

step six: and (3) credit decision making: and matching, integrating and processing the data information in the fifth step by the trust decision engine, calling the API interface service generated in the fourth step to obtain various risk assessment predicted values, matching different types of rules by the decision rule base to determine the risk performance of the customer, and giving personalized pricing.

Step seven: and (3) feeding back a credit result: and feeding back the personalized pricing obtained in the sixth step to the client, and displaying the personalized pricing through a front-end interactive tool.

The defects existing in the prior art are as follows:

1. enterprises only quantize credit earnings from the risk perspective of customers, cannot meet other scenes and dimensions, and cannot perform fine operation on the customers;

2. enterprises rely on risk perspectives to quantify credit earnings, and pricing strategy adjustment from self business perspectives is difficult to carry out.

Disclosure of Invention

Aiming at the problems that in the prior art, the fine operation of a client is difficult to realize only by quantifying the credit earnings from the risk perspective, and the pricing strategy is difficult to adjust from the business perspective by quantifying the credit earnings from the risk perspective by some small financial institutions, the invention provides a method capable of pricing credit finely, which aims to solve the problems that: the credit client cost and income related characteristic behaviors are quantized by combining a big data technology and a machine learning or deep learning technology, so that expected income prediction of the credit client is more accurate, fine operation of a credit product is realized, more dimensionality quantization indexes are provided for product operation, and the capacity of a financial enterprise for coping with complex market environments is improved.

In order to achieve the purpose, the invention adopts the following technical scheme:

a method for realizing decision evaluation based on multiple dimensions comprises the following steps:

step A: collecting risk characteristic data and income characteristic data of a client and storing the risk characteristic data and the income characteristic data into a big data platform;

and B: b, summarizing and sorting all the client data collected in the step A, and then processing;

and C: cleaning the processed data, confirming various target variables and total quantity characteristics, constructing models corresponding to different target variables, and performing model training and parameter optimization to obtain a trained model file;

step D: packaging the model file into API interface service, wherein the input of the API interface service is the characteristic required by the model, and the output is the result of model prediction;

step E: when a client applies for credit authorization in the front-end interactive tool, the acquisition of data required by the credit authorization of the client is completed in the service process;

step F, matching, integrating and processing the information acquired in the step E, wherein the processing logic corresponds to the processing logic in the step B one by one, then the processed data is served through the API interface generated in the step D to obtain output values corresponding to different models, different types of decision rules are matched according to the output values, the customer group division of the customer is determined, the personalized pricing of the customer is obtained, and the expected income of the customer is calculated;

step G: and displaying the personalized pricing through a front-end interactive tool and feeding back the personalized pricing to a client.

Further, in step a: the risk characteristic data and the income characteristic data of the customers comprise credit investigation data, third-party data, transaction data and customer behavior data of the customers, wherein the credit investigation data and the third-party data are from an external system, the transaction data are from enterprise data, and the customer behavior data are collected based on a front-end buried point tool.

Further, in step B: and matching and aligning all data related to the client according to the unique ID of the client, then counting the data, and finally performing basic processing.

Further, step C includes:

step C1: cleaning the processed data, confirming different target variables and total quantity characteristics, and selecting risk characteristics and income characteristics as different target variables respectively, wherein the data cleaning comprises data duplication removal and bad sample cleaning work, and the bad samples are data with missing values larger than 50%;

step C2: carrying out barrel separation or normalization on the continuity characteristics, calculating IG and chi-square values to carry out characteristic importance evaluation, and carrying out characteristic dimension screening in cooperation with statistical indexes;

step C3: dividing all the screened samples into N folds, and selecting corresponding models;

step C4: and carrying out model training through the selected model, and carrying out parameter tuning to obtain a trained model file.

The model trained by the method is not only one, but also can be modeled according to different requirements.

Further, step E includes: in the service process of applying for credit authorization, a client acquires data required by the credit authorization of the client, wherein the data acquisition comprises the following steps: basic information input by a client, client behavior data, information matched by calling a third-party service according to the unique ID of the client and transaction data matched with the interior of the enterprise according to the unique ID of the client.

Further, step F includes: the output values are risk predicted values and income predicted values corresponding to different models, and the obtained predicted value sets are matched through a decision rule base to determine the customer group division of customers; and meanwhile, the method carries out 'income calculation' for the client, and quantifies the expected income of the client.

The predicted values are essentially the probabilities of generating corresponding behavior characteristics by the user, the behavior characteristics are directly hooked with profits, costs and the like (such as withdrawal probabilities, borrowing probabilities and the like), and according to the probability of behavior occurrence and the generated profits, the calculation of the expected profits can be refined for a single customer, and the expected value of each customer is effectively quantized.

Further, the risk prediction value and the income prediction value include a withdrawal rate, an amount usage rate, a borrowing probability, a loss probability and a clearing probability in advance, and the income calculation specifically includes: customer expected revenue = (weight 1 × withdrawal rate) × (weight 2 × credit usage) × (weight 3 × borrowing probability) × (weight 2 × credit usage) × (weight 4 × loss probability) × (operation cost: - (weight 5 × early settlement probability) × (amount of settlement); the weight is the accuracy of the default use predicted value, and the adjustment of the actual application scene is supported.

According to the credit profit assessment method, a set of standardized and quantifiable credit customer profit index system is established based on the machine learning technology or the deep learning technology through characteristic behavior indexes, such as customer withdrawal rate, limit utilization rate, borrowing period, borrowing probability, early clearing probability, customer loss probability and the like, which are closely related to the credit profit of the user, so that the credit profit assessment in three aspects of dimensionalities, such as profit, cost loss, responsiveness and the like, is realized.

Furthermore, the decision rule base is configured through an expert background, the matched rules are accumulated corresponding to the expected profits of all the customers in the customer group, and the effect of the matched rules is quantized in return.

By adopting the optimal scheme, the invention quantifies the effect of the matching rule and can clearly and intuitively adjust the rule from the service perspective.

In summary, due to the adoption of the technical scheme, the invention has the beneficial effects that:

the credit return assessment method and the credit return assessment system combine a big data technology and a machine learning or deep learning technology to realize the quantification of the cost and return related characteristic behaviors of the credit customer, so that the expected return prediction of the credit customer is more accurate, the fine operation of a credit product is realized, and the credit return assessment on three aspects of return, cost loss, responsiveness and the like is realized.

And secondly, quantifying the guest group division rules by accumulating all income predicted values of the customers corresponding to the guest group division rules, thereby realizing the quantified estimation of the rule effect.

Drawings

FIG. 1 is a schematic flow diagram of the prior art of the present invention;

FIG. 2 is a schematic diagram of a customer quantization index system according to an embodiment of the present invention;

FIG. 3 is a flow chart illustrating an embodiment of the present invention.

Detailed Description

All of the features disclosed in this specification, or all of the steps in any method or process so disclosed, may be combined in any combination, except combinations of features and/or steps that are mutually exclusive.

The noun explains:

machine learning: machine learning is a data analysis technique that lets computers perform the activities inherent to humans and animals: learning from experience. Machine learning algorithms use computational methods to "learn" information directly from data, without relying on a predetermined equation model. These algorithms can adapt to improve performance as the number of samples available for learning increases.

Deep learning: deep learning refers to a multi-layered artificial neural network and a method of training it. One layer of neural network takes a large number of matrix numbers as input, weights are taken through a nonlinear activation method, and another data set is generated as output. The neural network brain is formed by connecting a plurality of layers of tissues together through proper matrix quantity, and the neural network brain is accurately and complexly processed just like a labeling picture of an object recognized by people.

Characteristic engineering: (Feature Engineering) it is a process of transforming raw data into features that better express the nature of the problem, so that applying these features to a prediction model can improve the accuracy of model prediction on invisible data. The feature engineering is to find features that have obvious influence on the dependent variable y, usually called independent variable x as features, and the purpose of feature engineering is to find important features. How to decompose and aggregate the original data to better express the nature of the problem is the purpose of feature engineering.

Customer breach probability PD: in the credit risk management of commercial banks, the default probability refers to the possibility that a borrower cannot pay back the original information of a bank loan or perform related obligations according to contract requirements within a certain period of time in the future. The default probability is the basis for calculating the expected loss of the loan, the pricing of the loan and the credit combination management, so how to accurately and effectively calculate the default probability is very important for the credit risk management of the commercial bank.

Debt loss rate LGD: the amount of loss that the debtor will cause to the creditor upon default, i.e. the severity of the loss. From a loan recovery perspective, LGD determines the extent of loan recovery, i.e., LGD = 1-recovery.

The invention will be further described with reference to the accompanying drawings and specific embodiments.

As shown in fig. 2 and 3, a method for implementing decision evaluation based on multi-dimension includes firstly constructing a set of single-customer quantitative index system highly related to profit and cost, which can be divided into profit, cost loss and responsiveness in total;

the income dimension comprises a client withdrawal rate, an amount utilization rate, a withdrawal period predicted value and a borrowing probability, and the characteristic behavior quantitative dimension of each client can directly reflect the income condition obtained by the enterprise from the client.

Cost loss dimensions reflect single customer characteristic behaviors that may cause direct losses to the enterprise, including: the loan client early settlement probability, the client attrition probability and the suspected capital operation probability. By quantifying these metrics, the loss caused by a single customer can be predicted.

The responsiveness dimension includes natural flow probability, interest rate sensitivity, and deep sleep customer survival promotion probability. The quantitative indexes can effectively evaluate marketing activities or service activities, and the higher the effective degree of customers is, the greater the business effect can be generated.

Based on the constructed income quantification system, the behavior characteristics of a single customer are predicted by combining user behavior data, transaction data and credit investigation data through a machine learning technology or a deep learning technology, and the predicted value is the corresponding quantification index.

in the step A: the risk characteristic data and the income characteristic data of the customers comprise credit investigation data, third-party data, transaction data and customer behavior data of the customers, wherein the credit investigation data and the third-party data are from an external system, the transaction data are from enterprise data, and the customer behavior data are collected based on a front-end buried point tool.

And B: collecting and sorting all the collected data, mainly comprising all the data related to the user, and matching and aligning according to the unique ID of the user; and (3) carrying out statistical processing on the user behavior data, wherein the functional pages related to the main business process are counted as the following indexes: last 3 days visit, last week visit, last month visit, no visit record days. Finally, some basic processing includes: null values, outlier normalization, code value unified translation, and the like.

the step C is specifically as follows: firstly, carrying out work such as duplicate removal and bad sample removal (the missing value is more than 50 percent) on data, and simultaneously confirming target variable and total quantity characteristics; then, carrying out characteristic engineering: and (3) calculating IG and chi-square values to evaluate the importance of the features by barrel division or normalization of continuous features, and performing feature dimension screening by matching with statistical indexes such as loss rate and the like. Then, the full sample is divided into N folds (i.e. usually 5-fold or 10-fold cross validation set is used), and a model (the model that can be tested includes GBDT, Xgboost, LightGBM, deep fm) is selected and then model training and parameter tuning are performed, and the tuning mode depends on the specific model, so that a trained model file is obtained finally.

The model trained by the invention is not only one, but also modeled by following different requirements, and the models mainly or jointly depict the client default probability PD and the debt loss rate LGD of the user. The modeling modes of the profit model training sub-process and the wind control model training sub-process are also consistent, and only in the selection of target variables, the risk characteristics (such as whether a user is overdue or not and whether bad account behaviors exist or not) are changed into the mode that profit-related data are used as the target variables (such as whether a client withdraws money or not, whether the client borrows the money or not and the like).

Step D: and packaging the model file into API interface service, wherein the API interface service can be developed based on Java or Python voice and provides an API interface for the outside, the input of the API interface service is the characteristic required by the model, and the output is the result of model prediction.

Step E: the client applies for the credit: the method is characterized in that a complete process and corresponding functions of client credit application and authorization are realized through Web service, public numbers or applets, mobile phone App and other front-end interactive tools, service is generated for the client through drainage, and then data required by client credit authorization are collected in the whole service process. The data acquisition comprises the following steps: basic information input by a user, user behavior data, relevant information matched by calling a third-party service according to the unique ID of the user, transaction data matched inside an enterprise according to the unique ID of the user and the like.

the step F comprises the following steps: the output values are risk predicted values and income predicted values corresponding to different models, and the obtained predicted value sets are matched through a decision rule base to determine the customer group division of customers; and meanwhile, the method carries out 'income calculation' for the client, and quantifies the expected income of the client. The nature of the predicted values is the probability of generating corresponding behavior characteristics by the user, the behavior characteristics are directly hooked with income, cost and the like (such as withdrawal probability, borrowing probability and the like), and according to the probability of behavior occurrence and the generated income, the calculation of the expected income can be refined for a single customer, and the expected value of each customer can be effectively quantized.

The risk prediction value and the income prediction value comprise a withdrawal rate, an amount utilization rate, a withdrawal period prediction value, a borrowing probability, a loss probability and a clearing probability in advance, and the income calculation specifically comprises the following steps: customer expected income = (weight 1 × withdrawal rate) × (weight 2 × quota utilization rate) × (weight 3 × withdrawal cycle predicted value × interest rate) + (weight 4 × borrowing probability) × (weight 2 × quota utilization rate) × (weight 3 × withdrawal cycle predicted value × interest rate) - (weight 5 loss probability) × operating cost- (weight 6 × early settlement probability) × settlement quota; each predicted value has an adjustable weight, and the weight can be used as a default for the accuracy of the predicted value and can be finely adjusted according to the actual exhibition situation on the basis.

Meanwhile, the precision of expected income of customers can be improved by continuously enriching similar related prediction probabilities, such as 'suspected fund operation probability' in an index system diagram, wherein the suspected fund operation probability refers to borrowing only in a free time and repayment after the free time, namely a so-called wool customer, and the cost loss is essentially caused; then we can calculate the added loss cost by adding the following formula.

New loss cost calculation = weight 8 suspected capital operation probability amount specific activity free period interest rate

And the expected accuracy is improved by incorporating the calculation formula into the overall calculation of the expected income of the client. The circulation of the process enriches the calculation of cost and income continuously, and is the core work of subsequent continuous optimization.

The decision rule base is a combination of a plurality of index screening conditions, for example (the predicted value A is greater than 0.7 and the predicted value B is between 0.2 and 0.3), the decision rule base is configured through an expert background, the matched rules are accumulated corresponding to the expected profits of all customers in the customer group, and the effect of the matched rules is quantized in turn. The effect of the matching rule is quantized, and the rule can be adjusted clearly and intuitively from the service perspective. For example: some rules cover a large number of customer groups, the customer responsiveness is high, but the profit amount is small, the rules are suitable for the business scene of impulse, and other rules cover a large number of customer groups, but the accumulated profit is large, so that the business can be achieved in real time when the KPI is triggered at the end of a month. Therefore, based on the quantized customer value and the quantization rule effect, flexible and quantifiable pricing strategies are realized.

According to the technical scheme, the big data technology and the machine learning or deep learning technology are combined, so that the cost and income related characteristic behaviors of the credit customer are quantized, the expected income prediction of the credit customer is more accurate, and the fine operation of a credit product is realized; meanwhile, accurate calculation of expected income of credit customers is realized through multi-dimension and systematized single-customer income quantification, the value of a customer group division rule (or strategy) can be further evaluated based on the calculation results, quantifiable and agile pricing strategy adjustment is realized from a business view, the richness, effectiveness and controllability of enterprises on the pricing strategy are greatly improved, and the credit products have core competitiveness for coping with the market responsible environment.

The above are merely representative of the embodiments of the present invention in many specific applications, and the scope of the present invention is not limited in any way. All the technical solutions formed by the transformation or the equivalent substitution fall within the protection scope of the present invention.

Claims

1. A method for realizing decision evaluation based on multi-dimension is characterized by comprising the following steps:

and C: cleaning processed data, including removing weight and removing bad samples, simultaneously confirming target variables and total quantity characteristics, then performing characteristic engineering, performing barrel separation or normalization on continuous characteristics, performing characteristic importance evaluation by calculating IG and chi-square values, performing characteristic dimension screening by matching with statistical indexes, wherein the statistical indexes comprise deletion rates, then dividing the total quantity samples into N folds, selecting corresponding wind control models and income models to perform model training and parameter optimization, and finally obtaining trained model files;

2. The method for implementing decision evaluation based on multiple dimensions as claimed in claim 1, wherein in step a: the risk characteristic data and the income characteristic data of the customers comprise credit investigation data, third-party data, transaction data and customer behavior data of the customers, wherein the credit investigation data and the third-party data are from an external system, the transaction data are from enterprise data, and the customer behavior data are collected based on a front-end buried point tool.

3. The method for implementing decision evaluation based on multiple dimensions as claimed in claim 1, wherein in step B: matching and aligning all data related to the client according to the unique ID of the client, then counting the data, and finally performing basic processing, wherein the basic processing comprises the following steps: null value, abnormal value standardization and code value unified transformation.

4. The method for implementing decision evaluation based on multiple dimensions as claimed in claim 1, wherein step C comprises:

step C1: cleaning the processed data, and simultaneously confirming different target variables and total quantity characteristics, wherein the target variables of the wind control model comprise that a client has no overdue deposit and no bad account deposit, the target variables of the income model comprise that the client has no withdrawal and no double borrowing, the risk characteristics and the income characteristics are respectively selected as different target variables, the data cleaning comprises data duplication removal and bad sample cleaning work, and the bad sample is data with a missing value of more than 50%;

step C3: dividing all the screened samples into N folds by using a 5-fold or 10-fold cross validation set, and selecting corresponding models;

5. The method for implementing decision evaluation based on multiple dimensions as claimed in claim 1, wherein step E comprises: in the service process of applying for credit authorization, a client acquires data required by the credit authorization of the client, wherein the data acquisition comprises the following steps: basic information input by a client, client behavior data, information matched by calling a third-party service according to the unique ID of the client and transaction data matched with the interior of the enterprise according to the unique ID of the client.

6. The method of claim 4, wherein step F comprises: the output values are risk predicted values and income predicted values corresponding to different models, and the obtained predicted value sets are matched through a decision rule base to determine the customer group division of customers; and meanwhile, the method carries out 'income calculation' for the client, and quantifies the expected income of the client.

7. The method of claim 6, wherein the method comprises: the risk prediction value and the income prediction value comprise a withdrawal rate, an amount utilization rate, a borrowing probability, a loss probability and a clearing probability in advance, and the income calculation specifically comprises the following steps: the customer expected income = (weight 1 is the withdrawal rate) ((weight 2 is the amount of use) ×) amount (weight 3 is the withdrawal period predicted value is the interest rate) + (weight 4 is the probability of borrowing) amount (weight 2 is the amount of use) × (weight 3 is the withdrawal period predicted value is the interest rate) - (weight 5 is the loss probability) operation cost- (weight 6 is the probability of early ending) ending amount, the weight is the accuracy rate of the default use predicted value, and the adjustment of the actual application scene is supported; and the accuracy of the expected income of the client is improved by continuously enriching the prediction probability.

8. The method of claim 7, wherein the method comprises: calculating the loss cost by adding the suspected capital operation probability:

new loss cost calculation = weight 8 suspected capital operational probability quota specific activity free period interest rate.

9. The method of claim 6, wherein the method comprises: and the decision rule base is configured through an expert background, the matched rules are accumulated corresponding to the expected profits of all the customers in the customer group, and the effect of the matched rules is quantized in turn.