Shulou(Shulou.com)06/03 Report--

For more information on big data's analysis and modeling, please follow the official account "bigdatamodeling".

BY pebble

1. Vintage

Vintage comes from wine-making, and the quality of wine varies according to the year in which the grapes are grown and the climate. Vintage analysis refers to the evaluation of the changes in the quality of wines of different years with the passage of time, and after a certain year, the quality of wine tends to be stable. As shown in the picture below, the quality of the wine in 2000 is the best. After about 5 years of cellar, the quality of the wine will tend to be stable.

With reference to wine Vintage analysis, the field of credit analysis can not only use it to evaluate the time required for customers to be fully exposed, that is, maturity, but also use it to analyze the differences of risk control strategies in different periods.

The quality of a customer needs to go through several repayment cycles before it can be fully exposed. if the performance period is short, it is possible to define a not-so-bad customer as bad or a very bad customer as good. For example, a customer will be overdue after overdue in the first few cycles, and the subsequent cycle will no longer be overdue, and if the performance period is short, the customer will be defined as bad; another example is that a customer has been making normal repayments at the beginning, but there is a serious overdue period later, and the customer will be defined as good if the performance period is short. In addition, we can analyze the differences of risk control strategies and the changes of macro situation by comparing the overdue performance of loans at different times in the same cycle.

The following figure shows the Vintage chart and Vintage data sheet of M4+ (that is, more than 91 days overdue) of loans from April 2017 to December 2018. The overdue status counted in this chart is the overdue status at the end of the month (historical overdue status can also be used in some cases. If a customer has overdue M4 +, the customer will be counted as M4 + overdue customer every month in the future), and the amount of overdue rate will be calculated (the number of pens can also be used). As can be seen from the chart, the M4 + of loans in different months tends to stabilize after 9 cycles (9 months), that is to say, the maturity period is in 9 cycles.

It can also be seen from the chart that the asset quality continues to improve, and the overdue rate from April to June 2017 is relatively high, and the overdue rate has declined to a large extent since July 2017, which may be due to the continuous optimization of risk control strategies in the past few months.

Second, rolling rate

The Vintage mentioned above can be used to analyze the trend of customer performance, stable time, etc., the definition of customer good or bad degree is not involved, but the customer good or bad degree can be defined by rolling rate analysis.

Roll rate analysis is the development and change from the worst state of a period of time before a certain observation point (called the observation period) to the worst state of the period after the observation point (called the performance period), as shown in the following figure.

At the observation point on June 30, 2018, 10,000 customers were selected to calculate the maximum overdue changes from the observation period to the performance period, as shown in the table below.

As can be seen from the table:

1. 96% of normal customers will remain normal in the next 6 months.

2. 81% of the customers who are overdue will return to normal in the future, that is, the recovery rate will be 81%, and 7% will worsen.

3. For customers who are overdue for the second period, the good rate is 23%, and 39% will worsen.

4. For customers who are overdue for the third period, the good rate is 14%, and 61% will worsen.

5. For customers with 4 or more overdue periods, the good rate is only 4%, and 82% will continue to do so.

When defining bad customers, bad customers should be defined as bad as possible to distinguish them from good customers as much as possible, so overdue periods 4 or more can be defined as bad, while overdue periods 2 and 3 are defined as moderately bad. Overdue period 1 is mildly bad. Of course, when developing models or rules, we should also combine the sample size. If M1 accounts for a relatively high proportion, M1 can be classified as a good customer, M4 + as a bad customer, M2 and M3 as grayscale customers, and removed from the sample; if the proportion of M1 is very low, M1, M2 and M3 can also be classified as grayscale customers and removed from the sample.

III. Definition of Y variable

The Y variable is the customer good or bad label variable, which has been mentioned in the previous Vintage analysis and roll rate analysis. Here is a summary. Y variable should be defined by combining roll rate analysis and Vintage analysis, roll rate analysis is used to define customer quality, and Vintage analysis is used to set an appropriate performance period. Defining the Y variable can be divided into the following steps:

1. Conduct roll rate analysis to define bad customers, such as M4 + as mentioned above.

2. Then count out the Vintage data table and Vintage chart of M4 +, and find out the mature period.

3. The samples whose performance period is larger than the mature period can be used for modeling, and the samples whose performance period is less than the mature period can not accurately define the Y variable and abandon it temporarily.

Of course, in practical application, the definition of Y variable can be flexible according to business needs. For example, shortly after the start of the business, the performance period is short, and the Y variable cannot be defined according to the above method. Customers who are more than 10 days, 15 days or 30 days overdue can be defined as bad, and then the Y variable and model can be modified over time.

4. Mobility

Mobility analysis and roll rate analysis are similar to analyzing the development and changes of customers from one state to another, but the difference is that the roll rate focuses on analyzing the change of the overdue degree of the customer, so it is necessary to set a relatively long observation period and realization period when doing the roll rate analysis, while the mobility analysis focuses on analyzing the development and change path of the customer state, such as M0 situation, M 1, M 1, M 2, and M 2, etc.

First, statistics are made on the asset distribution of each overdue state from June 2018 to December 2018, as shown in the table below. The recovery rate in the table refers to the assumption that M7 assets are sold at 10% price at the end of each month, that is, 10% of M7 assets are recovered.

The monthly mobility is calculated below. For example, at the end of July 2018, the mobility of M0~M1 is 41110122. By analogy, the mobility of the following table is obtained. Here, the mobility is calculated on the basis of month as the time granularity. At present, the mobility can also be calculated as fine as the day granularity, and then the average monthly mobility can be calculated, but the finer granularity requires that the asset scale is larger. If the asset scale is smaller, there will be a larger error. It is worth noting that the migration rate of M3~M4 in the table is relatively high, which should be more than 90 days overdue, after the gold collection period, so the overdue deterioration is higher, but the mobility rate of M4~M5 has suddenly decreased a lot, which may be due to the effect of external collection.

The loss rate of each overdue state is calculated according to the migration path. The ratio of M0 to M7 is defined as the loss rate equal to 16.06% "29.27%" 42.28% "81.71%" 52.75% "82.51%" 86.05% "0.61%. Because M7 will be recovered at a price of 10%, the net loss rate is equal to 0.61%" 90% "0.55%, and so on, the loss rate of assets in each overdue state is obtained, as shown in the table below.

When we talked about the definition of Y variable, we define the degree of good or bad according to the rolling rate analysis. In fact, we can also analyze the degree of good or bad according to the loss rate of mobility, and we can also combine the loss rate with the rate of return.

The expected asset loss in December 2018 is calculated based on the asset loss rate, which is calculated by multiplying the assets of M0~M6 by the corresponding net loss rate in December 2018, and then summing up, which is equal to 3671628090.55% "47230430" 3.41% "14848678" 11.64% "6011499" 27.54% "4614038" 33.70% "2326454" 63.89% "1586471" 77.44% "11273470". Therefore, in December 2018, the amount of provision was 11273470, accounting for 11273470 of the total assets, accounting for 443823814 million 2.54%, that is, the provision rate was 2.54%.

Note: the data in this paper is not the real data, but the data generated to illustrate the problem.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.