CN112819641A

CN112819641A - Intelligent pricing system, method, equipment and storage medium for group insurance

Info

Publication number: CN112819641A
Application number: CN202011585015.0A
Authority: CN
Inventors: 王玮; 徐勤燕; 顾佳盛; 钟严堃
Original assignee: Shanghai Data Center of China Life Insurance Co Ltd
Current assignee: Shanghai Data Center of China Life Insurance Co Ltd
Priority date: 2020-12-28
Filing date: 2020-12-28
Publication date: 2021-05-18

Abstract

The invention relates to a system, a method, equipment and a storage medium for intelligent pricing of a group insurance, wherein the pricing system comprises: the foreground page module is used for inputting new order insurance information and client information; the intelligent pricing model module is used for predicting the probability of the client insuring a certain dangerous type in seconds according to the input information of the client and the dangerous type and further calculating the predicted claims and pricing; the web interface module is used as an intermediary to enable a user to trigger model prediction by sending an HTTP request calling interface which accords with data input rules; and the background data storage module is used for storing all relevant service data in the big data platform. The invention creates a set of efficient and intelligent risk group pricing models, reduces the difference risk and the experience approval risk in risk group business, reduces the dependence on people and improves the timeliness and the accuracy of pricing; and the deep learning method is combined with the big data technology to be applied to the insurance business, and a new idea of applying the artificial intelligence technology to the insurance business is developed.

Description

Intelligent pricing system, method, equipment and storage medium for group insurance

Technical Field

The invention relates to the technical field of hedging pricing, in particular to an intelligent hedging pricing system, method, equipment and storage medium.

Background

The group insurance business is the focus of the competition of the big insurance companies in the insurance market and is the cost-creating profit point of the basic insurance company, and the pricing of the group insurance business is an important component of the group insurance product. Whether the pricing is reasonable or not determines the benefits of the insurance company business operation and whether the applicant can purchase the maximum insurance guarantee at a reasonable price or not.

The group insurance intelligent pricing system is a system for calculating expected claims and expected premium by predicting the rate at the second level through a deep neural network intelligent model according to customer information and purchased product information, and supporting the evaluation of new order pricing of group insurance business. The system in the technical scheme of the invention is based on company business reality, and mainly aims at the problems that a group insurance expert is relied on in the traditional group insurance business, the risk evaluation process is long, and the pricing result is slow in feedback timeliness; the rate calculated by a tabulation method, an empirical method and the like usually ignores the difference risk of customers, so that the price of a company product is lack of elasticity, and the company product is lack of competitiveness in a fierce group insurance market.

Disclosure of Invention

The present invention is directed to a system, method, device and storage medium for intelligently pricing a group risk, which overcomes the above-mentioned drawbacks of the prior art.

The purpose of the invention can be realized by the following technical scheme:

the utility model provides a quest intelligence pricing system, this system includes foreground page module, intelligent pricing model module, web interface module and backstage data storage module, wherein:

the foreground page module is used for inputting new order insurance information and customer information;

the intelligent pricing model module is used for predicting the danger probability of a client for putting a certain dangerous type according to the input information of the client and the dangerous type in seconds, and further calculating predicted claims and pricing;

the web interface module is used as an intermediary to enable a user to trigger model prediction by sending an HTTP request calling interface which accords with data input rules;

and the background data storage module is used for storing all relevant service data in the big data platform.

Further, the intelligent pricing model module adopts an intelligent pricing regression model which is formed by fusing a plurality of neural networks through a conticatenate layer, wherein each neural network is composed of a Dense layer and a Dropout layer, the activation function of the intelligent pricing regression model adopts a ReLU function, the loss function adopts an MSLE function, the optimizer adopts ADAM, and the minimum batch min-batch is set to be 128.

Further, the web interface module is realized by adopting a flash frame of Python.

Further, the model input of the intelligent pricing regression model in the intelligent pricing model module comprises the name of a company, the industry, the occupation, the subsidiary of the workplace, the quota, the premium, the total number of people, the average age and the male proportion; the model outputs include the exposure rate, the mean claim, and the average number of hospitalizations.

The invention also provides a group risk intelligent pricing method, which is implemented by adopting the group risk intelligent pricing system and comprises the following steps:

step 1: after the user inputs new single-delivery insurance information and client information in the foreground page module, the web interface module is used as a medium to enable the user to call an interface by sending an HTTP request meeting data input rules, and model prediction is triggered;

step 2: and the intelligent pricing model module predicts the probability of the client insuring a certain dangerous type in seconds according to the input client and dangerous type information, further calculates the predicted claims and pricing, and returns to the foreground page module to be displayed to the user.

Further, the intelligent pricing model module in the intelligent pricing method for the risk adopts an intelligent pricing regression model formed by fusing a plurality of neural networks through a concatenate layer, wherein each neural network is composed of a Dense layer and a Dropout layer, the activation function of the intelligent pricing regression model adopts a ReLU function, the loss function adopts an MSLE function, the optimizer adopts ADAM, and the minimum batch min-batch is set to be 128.

Further, the web interface module in the intelligent risk-fighting pricing method adopts a web interface module realized by a flash frame of Python.

Further, the model input of the intelligent pricing regression model in the intelligent pricing model module in the group insurance intelligent pricing method comprises company name, industry, occupation, branch company of the workplace, premium, total number, average age and male proportion; the model outputs include the exposure rate, the mean claims, and the average days in hospital.

The invention also provides a terminal device, which comprises a memory, a processor and a computer program stored in the memory and capable of running on the processor, wherein the processor realizes the steps of the intelligent pricing method for the risk when executing the computer program.

The invention also provides a computer-readable storage medium, which stores a computer program that, when executed by a processor, implements the steps of the intelligent pricing method for a risk.

Compared with the prior art, the invention has the following advantages:

(1) the invention develops a set of risk intelligent pricing regression model, gives consideration to the difference risk and the timeliness in the risk service, and the model can predict the risk probability of the client for insuring a certain risk class in a second level by inputting the information of the client and the risk class so as to predict the benefits and pricing.

(2) According to the invention, a set of efficient and intelligent risk pricing model is created, so that the differential risk and the experience approval risk in the risk business are reduced, the dependence on people is reduced, and the timeliness and the accuracy of pricing are improved; and the deep learning method is combined with the big data technology to be applied to the insurance service, and a new idea is opened up for applying the artificial intelligence technology to the insurance service.

(3) The traditional risk pricing calculation method is complex, and has different information influencing product pricing and different pricing rate calculation modes for different regions, different risk varieties and different responsibilities. The created intelligent pricing model for the risk combines the difference risk and the timeliness in the risk service, and the model can predict the rate of the client for applying a certain risk class in second level to predict the expected pricing as long as new client information is input. Compared with the traditional group insurance pricing approval, the risk can be efficiently and accurately evaluated according to different characteristics of clients and products, and the dependence on group insurance experts is reduced.

(4) The machine learning technology is the core technology of artificial intelligence, and the essence of the machine learning technology is the research on a computer algorithm which can be automatically improved through experience, and common machine learning algorithms comprise a decision tree, a random forest, an SVM, naive Bayes, a neural network and the like. The neural network is inspired by human brain and is an interconnection relation between neurons. In recent years, the processing capacity of computers is increased by tens of millions, and the rise of cloud computing pushes the heat of neural networks. Deep Learning (Deep Learning) is a new field of machine Learning, is essentially a multilayer neural network, and has certain superiority in aspects of feature engineering, model fusion, incremental training data and the like compared with other machine Learning methods.

At present, no research for practicing artificial intelligence technology in group insurance pricing exists in China. In deep learning, the author refers to a scenario of machine learning task, in which the amount of a claim is predicted by regression method for setting insurance fee. Therefore, after a plurality of machine learning algorithms are tried, the invention innovatively applies deep learning and combines big data technology to change the pricing mode of the group insurance supported by insurance actuarial and business experience.

Drawings

Fig. 1 is a schematic diagram of a premium and a headcount distribution in an embodiment of the present invention, in which fig. 1(a) is a schematic diagram of the premium distribution and fig. 1(b) is a schematic diagram of the headcount distribution;

FIG. 2 is a schematic diagram of the distribution of the total number of people and the amount of the reserves after log extraction in the embodiment of the invention, wherein FIG. 2(a) is a schematic diagram of the distribution of the reserves after log extraction, and FIG. 2(b) is a schematic diagram of the distribution of the total number of the people after log extraction;

FIG. 3 is a flowchart illustrating operation of the intelligent pricing model according to an embodiment of the invention;

FIG. 4 is a schematic structural diagram of a deep neural network regression model according to an embodiment of the present invention;

FIG. 5 is a flow chart of a web interface service in an embodiment of the invention.

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are some, not all, embodiments of the present invention. All other embodiments, which can be obtained by a person skilled in the art without any inventive step based on the embodiments of the present invention, shall fall within the scope of the present invention.

DETAILED DESCRIPTION OF EMBODIMENT (S) OF INVENTION

The group insurance intelligent pricing system is a system for calculating expected benefits and expected premium by predicting the rate in second level through a deep neural network intelligent model according to customer information and purchased product information, and supporting group insurance service evaluation new order pricing. The system mainly comprises a foreground page, an intelligent pricing model, a web interface, background data storage and the like. The foreground page is mainly responsible for inputting new order insurance information, client information and the like. The intelligent pricing model is an intelligent pricing regression model built in a mode of fusing a plurality of deep neural networks, and as long as information of a client and dangerous species is input, the model can predict the probability of the client insuring a certain dangerous species in a second grade so as to predict the benefits and pricing. Background data is stored on the big data platform. The WEB interface is realized through a flash frame of Python and is grounded on a large data platform. And the user triggers model prediction by sending an HTTP request calling interface which accords with the data input rule to obtain the intelligent pricing service of the group risk. The predicted data is immediately sent back to the foreground for presentation to the user.

The principle of the modeling process of the intelligent pricing model of the intelligent pricing system for the risk is as follows:

the modeling implementation steps comprise data acquisition, data cleaning, feature engineering, modeling, evaluation, tuning and release. Data cleaning comprises data processing technologies such as data distribution inspection, null value processing and abnormal value processing, and is a very important step in the early stage. The source data typically contains a variety of dirty data that affects the accuracy of model training and prediction, requiring data cleansing. The feature engineering comprises feature selection and dimension reduction, and finally a feature vector is formed. Modeling includes model selection, model training, and algorithm development. The evaluation and optimization comprise super-parameter optimization, model structure optimization, cross validation and the like. And (4) evaluating the model effect according to the tuning result, and achieving the ideal model effect after repeated feature adjustment, model fine tuning and parameter tuning possibly.

Data acquisition, data cleaning, feature engineering, modeling, evaluation, tuning and publishing

Data acquisition:

the training data of the group insurance intelligent pricing system totally tries three different extraction schemes.

The first scheme is as follows: the scheme groups and aggregates statistical characteristics and output according to dangerous species (responsibility, industry and occupation); model features (risk liability, industry, occupation, headcount, total number of claims, average age, male proportion), model outputs (risk rate, average claims, average days in hospital).

Scheme II: according to the scheme, data are extracted according to a single insured person, and the output of a model obtained by training represents the output of the insured person; model features (risk liability, industry, occupation, age, gender), model outputs (coefficient of venture, amount of claims, days of hospitalization).

The third scheme is as follows: the scheme groups and aggregates statistical characteristics and outputs according to (insurance unit, risk code, risk responsibility, industry, occupation and public driver structure number). Model features (risk category code, risk category responsibility, industry, occupation, division organization number, premium, headcount, average age, male proportion), model outputs (rate of occurrence, per-piece claims, average days of hospitalization).

The first scheme has the advantages of less characteristics, larger grouping aggregation granularity, only 3073 data quantity and less data quantity; the second scheme is that a single insured person extracts data, the data size is 2866 thousands, the model is not easy to converge, and the generalization capability is poor; and on the other hand, the grouping aggregation granularity is moderate, the data volume is 20 thousands, and the training requirement of the DNN deep learning model can be met. Therefore, the data is finally extracted by adopting the third scheme.

Data preprocessing:

in the model features extracted according to scheme three: the dangerous seed code, the dangerous seed responsibility, the industry, the occupation and the public driver structure number are discrete characteristics and need to be coded; the quota, the total number of people, the average age and the male proportion are continuous characteristics, and barrel separation, scale transformation and normalization processing are required.

1. Average age treatment: continuous feature discretization

The average age given by the business personnel in the inquiry stage is an estimated value, and 7 different types of age groups are obtained according to the average age of 19, 36,46,56,61 and 66 according to business experience, so that the continuous features can be converted into discrete features.

2. Discrete feature coding

One-hot coding is adopted for discrete features, and for each feature, if the feature has m possible values, the feature becomes a vector with the length of m after the unique hot coding, only 1 bit in the vector is 1, and the rest bits are 0.

The risk category codes have 24 categories, the risk category responsibility has 13 categories, the industry has 22 categories, the occupation has 88 categories, the branch organization number has 20 categories, and the average age group has 7 categories, so that the discrete feature codes are spliced by 6 one-hot vectors, and the total length is 174.

3. Continuous feature scale transformation

The distribution of the amount (amount) and the total number (num) is shown in FIG. 1(a) and FIG. 1 (b). The mean value of amount is 1060.57, the variance is 15957.59, the num mean is 140.65, and the variance is 3318.95. As can be seen, the range of the reserves and headcount is quite large, and therefore compression of larger data is required.

The number of the total (amount) and the total number (num) are logarithmized, and the distribution is shown in FIG. 2(a) and FIG. 2 (b). The mean value of log _ average is 2.06, the variance is 0.81, the mean value of log _ num is 1.36, and the variance is 0.68. Comparing fig. 1(a) and fig. 1(b), the data range is effectively compressed after log extraction, and the mean and variance are obviously improved.

4. Continuous feature normalization

The continuous features are log _ amount, log _ num and male proportion, different features have different distribution areas, and normalization processing is needed to enable the features to be in a relatively equal position. The project is normalized by Z-Score, and the formula is as follows:

intelligent pricing model

As shown in FIG. 3, the intelligent pricing model is an intelligent pricing regression model built in a mode of fusing a plurality of deep neural networks, and as long as information of customers and dangerous species is input, the model can predict the risk probability of the customers for insurance of a certain dangerous species in a second level so as to predict the claims and pricing. The intelligent pricing model can be incrementally optimally adjusted based on user usage feedback.

Concrete modeling method

Through test comparison of different machine learning algorithms, the invention finally selects a mode of fusing a plurality of deep neural networks to build a deep learning model. The model structure is shown in fig. 4, each neural network is composed of a sense layer (full connection layer), a Dropout layer, and the like, and finally, a plurality of neural networks are fused through a concatenate layer. The integrated deep learning model integrates the advantages of various network structures, and uses a Dropout layer to prevent the model from being over-fitted, so that the accuracy of model prediction is improved. The network structure is characterized in that the feature codes directly enter 5 sub-networks after being input, and the structures and initialization methods of the 5 sub-networks are different. Each subnetwork has a 4 to 6 layer structure and is eventually pooled to a terminal node. And constructing a loss function through the final node and the actual real value, optimizing the value of each layer of weight of the neural network by means of a neural network back propagation algorithm, and finally finishing the training of the deep neural network model through continuous optimization iteration.

Method for building deep learning model by fusing multiple deep neural networks

Brief introduction to the model

Model input discrete features: policy, coverage, industry, occupation, city, avg _ age; continuous features: log _ amount, log _ num, sex _ ratio. The code length of the discrete features is 174 bits, and the code length of the continuous features is 3 bits. Discrete features and continuous features are concatenated, for a total of 177 bits of code as input to the model.

The deep neural network regression model is shown in fig. 4 and is composed of a density layer, a Dropout layer and the like. The network structure is characterized in that the feature codes directly enter 4 sub-networks after being input, and the structures and initialization methods of the 4 sub-networks are different. ReLU is used as an activation function of the deep neural network, MSLE is used as a loss function, ADAM is used as an optimizer, and min-batch is set to be 128.

Model evaluation index

And MAE, RMSE and R-square are adopted as evaluation indexes of the deep learning regression model, wherein the MAE and the RMSE are used for calculating the deviation between a predicted value and a true value, and the smaller the two indexes are, the more accurate the prediction result is. R-square is used to measure the fit of the model, and the index is usually within the interval of (0,1), with closer to 1 indicating better regression fit of the model. Using a sample data test model to obtain four prediction indexes:

1. rate of occurrence of danger

Table 1 shows regression indexes of the risk, the value range [0, ∞ ] of the risk, and the R-square of the verification set can reach 0.512, so that the regression effect is good.

TABLE 1 regression index of risk ratio

2. Pay for the equal part

Table 2 shows regression indexes of the piece average payment, the value range [0, ∞ ] of the piece average payment, and the R-square of the verification set can reach 0.528, so that the regression effect is good.

Table 2 piece per payout regression index

3. Claims on average

Table 3 shows regression indexes of mean claims, the value range [0, ∞ ] of the mean claims, and the R-square of the verification set can reach 0.444, so that the regression effect is good.

Table 3 mean claims regression index

4. Average number of hospitalization days

Table 4 shows the regression index of the average number of hospitalization days, the range [0, ∞ ] of the average number of hospitalization days, and the R-square of the validation set is 0.300, and the regression effect is general.

TABLE 4 mean days to stay regression index

WEB interface

As shown in fig. 5, a lightweight WEB interface is implemented by a flash framework of Python, and is grounded on a large data platform. And the user triggers model prediction by sending an HTTP request calling interface which accords with the data input rule to obtain the intelligent pricing service of the risk. The predicted data is immediately sent back to the foreground for presentation to the user. Meanwhile, the WEB interface also has strong agility and stability, can be updated quickly according to the change of the model, the whole process does not exceed 1 minute, and the interface service is not interrupted.

Data processing and storage

The data is stored in a big data platform, and business data is imported into the HDFS through sqoop. Data washing was performed using hive, Impala.

The system inputs information:

company name, industry, occupation,

Division, premium, total number of people, average age, male proportion of the workplace

The system outputs 4 variables:

rate of occurrence, claims per part, average days of hospitalization

According to the four values, the expected compensation amount is calculated by combining with other input information such as the insurance amount, and therefore, an intelligent quotation is provided for the client.

The key points and the protection points of the invention are to develop a set of intelligent pricing regression model, and the model can predict the risk probability of the client for insurance of a certain risk class in second level by inputting the information of the client and the risk class so as to predict the claims and pricing. The model is built by adopting a mode of fusing a plurality of deep neural networks, each neural network is composed of a Dense layer (a full connection layer), a Dropout layer and the like, and the plurality of neural networks are fused through a concatemate layer. The integrated deep learning model integrates the advantages of various network structures, and the Dropout layer is used for preventing the model from being over-fitted, so that the accuracy of model prediction is improved. The network structure is characterized in that the feature codes directly enter 5 sub-networks after being input, and the structures and initialization methods of the 5 sub-networks are different. Each subnetwork has a 4 to 6 layer structure and is eventually pooled to a terminal node. The invention constructs the loss function through the final node and the actual real value, optimizes the weight value of each layer of the neural network by means of the neural network back propagation algorithm, and finally completes the training of the deep neural network model through continuous optimization iteration.

While the invention has been described with reference to specific embodiments, the invention is not limited thereto, and various equivalent modifications and substitutions can be easily made by those skilled in the art within the technical scope of the invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. The utility model provides a quest intelligence pricing system which characterized in that, this system includes foreground page module, intelligent pricing model module, web interface module and backstage data storage module, wherein:

the intelligent pricing model module is used for predicting the probability of the client insuring a certain dangerous type in seconds according to the input information of the client and the dangerous type, and further calculating predicted claims and pricing;

2. The intelligent pricing system for a group risk according to claim 1, wherein the intelligent pricing model module employs an intelligent pricing regression model fused by a plurality of neural networks through a conticatenate layer, wherein each of the neural networks is composed of a Dense layer and a Dropout layer, the activation function of the intelligent pricing regression model employs a ReLU function, the loss function employs a MSLE function, the optimizer employs ADAM, and the minimum batch min-batch is set to 128.

3. A system for intelligent pricing of hedging according to claim 1, characterized in that the web interface module is implemented by using the flash framework of Python.

4. A hedging intelligent pricing system according to claim 1, wherein the model inputs of the intelligent pricing regression model in the intelligent pricing model module include company name, industry, occupation, division of the workplace, premium, headcount, average age, and male proportion; the model outputs include the exposure rate, the mean claim, and the average number of hospitalizations.

5. A method for intelligent pricing of hedging, implemented with the system for intelligent pricing of hedging according to any of claims 1 to 4, characterized in that it comprises the following steps:

step 1: after the user inputs new single-delivery insurance information and client information in the foreground page module, the web interface module is used as a medium to enable the user to trigger model prediction by sending an HTTP request calling interface which accords with a data input rule;

step 2: and the intelligent pricing model module predicts the probability of the client insuring a certain dangerous type in seconds according to the input information of the client and the dangerous type, further calculates the predicted claims and pricing, and returns to the foreground page module to be displayed to the user.

6. The intelligent pricing method for a group risk according to claim 5, wherein the intelligent pricing model module employs an intelligent pricing regression model fused by a plurality of neural networks through a conticatenate layer, wherein each neural network is composed of a Dense layer and a Dropout layer, the activation function of the intelligent pricing regression model employs a ReLU function, the loss function employs a MSLE function, the optimizer employs ADAM, and the minimum batch min-batch is set to 128.

7. The intelligent pricing method for a group risk according to claim 5, wherein the web interface module is implemented by using a flash framework of Python.

8. The intelligent pricing method for a group risk according to claim 5, wherein the model input of the intelligent pricing regression model in the intelligent pricing model module comprises company name, industry, occupation, branch of the workplace, premium, headcount, average age and male proportion; the model outputs include the exposure rate, the mean claim, and the average number of hospitalizations.

9. A terminal device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor when executing the computer program implements the steps of the intelligent pricing method for hedging according to any of claims 5-8.

10. A computer-readable storage medium, in which a computer program is stored which, when being executed by a processor, carries out the steps of the intelligent pricing method for hedging according to any of the claims 5 to 8.