On December 8th, open this year's calendar, 2021 still has more than half a month to go, and the last month or less allows us to grasp the rest of the time tail and have a wonderful (fishing) December.
Tomato risk control recently began to broadcast the content of small micro risk control, recently communicated with the students of the planet, everyone is more interested in the small micro model, today to share with you two major models on the small micro risk control: revenue estimation model and overdue forecast model.
One. Revenue estimation model
1.1. Model Background
In the business scenario of the enterprise, we always hope to use the historical operation data of the enterprise to predict the future operation. For example, we are using the data of the past 2 to 3 years to predict the actual turnover in the next 1 year, so based on the needs of such scenarios, we have the content of the revenue estimation model.
In the revenue estimation model, we have been using two major types of algorithms, one is a linear regression model, the other is an ARMA model (time series analysis model), these two types of models estimate revenue is somewhat different, see the following:
1.2. Comparison of model scenarios:
For different trend forecasting of turnover estimates, the models used are different, and the method of actual testing is used to find out the most suitable method for the turnover forecast of the enterprise, and the ARIMA model (time series analysis) is more suitable for enterprises with large turnover fluctuations.
Linear regression model in the turnover trend is more obvious in the enterprise is more suitable, in the past 2 to 3 years of data to predict the actual turnover of the past 1 year, we use the actual data to verify the accuracy of the model, to select the appropriate model for the actual next 1 year turnover forecast.
In the estimation of actual small and micro data, the actual test down, we summarized the following rules:
30% of enterprises are suitable for time series models,
30% of enterprises are suitable for linear regression models,
The remaining 40% of businesses that are doing well are better suited to the moving average estimate.
After the prediction method is tested in the actual data after selection, more than 80% of the customer accuracy can be controlled within the error of plus or minus 20%, and 8 of the 10 customers are predicted to be 1 million, fluctuating up and down not more than 20%, not more than 120w, and no less than 80w. The actual data test, the accuracy of the test is in line with the business production needs.
1.3. The model involves fields:
1.4. Comparison of forecast results
The overall prediction error rate of the ARIMA model is less than the linear regression at each percentile value, which shows that the ARIMA model is more accurate in forecasting revenue;
ARIMA forecast accuracy: 92.3%;
Linear regression prediction accuracy: 89.4%;
Two. Overdue forecasting model
2.1. Model Background
The overdue probability predicts that the customer it targets is basically a non-bank or bank, and predicts what is the probability that after he has borrowed a sum of money, he will not be able to repay the loan in the end.
This applies a lot of data, such as the basic information of the enterprise, the number of years of establishment, the registered capital, the business judiciary, the tax declaration sales, all of which are integrated together, and through this information to predict the probability of whether its final loan will default, this is mainly a logistic regression
2.2. The output model of the model
Customer overdue probability[0-1]
For example, through the overdue model, it is estimated that a limited liability company in Guangdong has a loan overdue probability of 0.00051, and the overdue risk is low.
Through the estimation of the probability of overdue, we can determine the probability of overdue, and the overdue situation can still be divided into interval levels.
2.3. The model involves fields:
【Excerpt Field】
2.4. Comparison of forecast results
Combined with our previous development data, through some previous model data statistics, the model's outcome reference index is used in the prediction of the training sample:
The above model refers to the corporate tax data (INTERNAL TAX), and the scope of tax payment of enterprises is at the level of hundreds of millions of annual sales, and the product range can reach 1 million.
In addition to the above two common models on the small micro risk control model, there are also:
(3) Operating a fraud model
(4) Withdrawal rate (balance) forecast model
Among them, the business fraud model is based on the customer's invoicing information to predict the risk of false operation of customers; while the balance prediction model is based on the actual withdrawal of the historical application customer, combined with the basic information of the enterprise, industrial and commercial information, fiscal and tax information to predict the balance and withdrawal rate of the model, the remaining two parts of the follow-up opportunity to introduce to you.
Regarding the mentioned model of several large and small micro risk control, more detailed content attention: "The second phase of the small and micro risk control training camp", this training camp we updated the model of small and micro enterprises + enterprise credit related content:
In addition, the enterprise credit tax field involved in this article, this time there is also relevant content in the knowledge planet platform for everyone to learn, you can check it above.
~ Original article
...
end