Before we start the analysis, we need to understand the process of loan business, that is, all the links involved in each user from purchase to final repayment. Generally speaking, the process of online lending can be summarized as follows:
Through user behavior path analysis, we can get:
It is suitable for analyzing and monitoring the key links in product operation, finding the weak links, optimizing through user guidance or product iteration, and improving the transformation effect.
Select a subset to count the number of new users and old users who apply for loans and approve loans every day, and then calculate the lending rate of new users. Finally, the new user result table is spliced with the old user result table by using the merge function. The results are as follows:
Next, we need to know the number of old users every day to calculate the repayment rate of users. Here, the definition of an old user is: # The new user who borrowed the day before and continued to borrow the next day is an old user #. For the existing old users, we don't consider it for the time being, but only see whether the borrower who borrowed the previous day continues to borrow the next day. This loan is regarded as the repayment of the old user's loan, so we take the data of the first 29 days of the new user's loan pivot table+May 1 personnel composition.
Form a summary table of user paths and calculate the conversion rate of each node.
Calculate the transformation funnel and calculate the summary data:
The results are as follows:
From the results, there is a great decline from PV to UV. Under normal circumstances, an advertisement may attract about 10%-30% users, but the number of clicks is very large, so this result is caused.
On the whole, it is obvious that 4% of the registered people are relatively low, which means that users click a lot, but the registered people are very few. Does it mean that there is a problem with this channel? I can compare the conversion funnel analysis of different channels to see if the conversion rate of the whole industry is low, or if the single channel is low, or if the user base of this channel is not what we want.
In consumer finance companies, it is often found through daily monitoring that the one-time excess rate of a consumer loan product is gradually rising. We need to reduce the first pass rate to reduce the loss caused by the product, and the pass rate should not be too obvious.
Analysis goal: through data exploration and analysis, formulate strategies that can effectively reduce the first pass rate.
Analysis idea: Because the strategy we want to analyze will be used to judge whether the customer will be overdue when applying, the basic idea of strategy analysis is to restore the data of these customers with initial performance when applying (this restoration refers to extracting the data of customers in all dimensions when applying, the more the better), and then use these data to find out the variables that can distinguish good and bad customers and make strategies.
It can be concluded that the overall one-time pass rate is 30.76%.
Univariate analysis is used here. The main purpose of univariate analysis is to screen out the variables that distinguish good from bad, so as to make strategies. In the daily work of Xiaojin Company, there will be a team responsible for grabbing variables and calculating the data of processing variables. They are constantly acquiring and processing a lot of data that may be helpful to risk control, and providing it to our risk control team. Our risk control personnel need to find out the variables that can control the overdue risk from these thousands of variables, without wrongly rejecting many good customers.
The statistical results are as follows:
The statistical results are as follows:
After the variable analysis, at this time, we will select more effective variables, which involves an index to measure whether the variables are effective and improve the degree. Generally speaking, it is to measure the improvement effect of overall risk control after rejecting the worst customers. The higher the promotion degree, the more effectively the variables can distinguish good customers from bad customers, and the less likely it is to reject good customers by mistake. As follows, by arranging the promotion degrees of all variables in reverse order, it is found that the total number of personal credit inquiries and the promotion degree of customer credit rating are the highest, reaching 1.93 and 1.7 1 respectively.
Through the univariate analysis in the previous step, we screened out two variables with the highest improvement: "credit inquiry times" and "credit rating". If the customer with the worst of these two variables is rejected, the impact on the overall overdue will be. This effect means that we assume' the total number of credit inquiries >; After 3,265,438+03 customers = 265,438+0 all rejected, the overdue rate of other customers decreased compared with that before rejection. In the end, we got that the number of credit inquiries decreased by 3.4% and the credit rating decreased by 7.5%.
Users have a process of user behavior in the process of using products, and the performance of users may be different in different periods. The main purpose of group analysis is to analyze the changes of similar groups over time, and the core is to compare and analyze the behavior differences of users in different groups at the same time, so it is also called simultaneous group analysis.
Then, in the field of financial risk control, the most commonly used scenario is aging analysis, which is used to monitor the change of overdue rate of users. As shown in the figure below, the overdue rates of M2 and M3 are relatively high, and then the risk control strategy is adjusted, and then the group analysis is carried out to see if the strategy is effective.
The results are as follows:
The new field generated here, orderperiod is the month when the user placed the order, and cohortgroup generates the user group according to the earliest period when each user placed the order, which are different.
Next, group by user group and month field:
The results are as follows:
In the orderperiod field, we can see that the earliest consumption month corresponding to 20 19-0 1 is 2009-0 1 02,03,05, but the earliest consumption month corresponding to 2009-02,03,05 is relative.
Get:
Therefore:
Note that each column in the above figure represents the earliest consumption group in the current month, and cohortperiod represents the situation of the earliest consumption group in 1, February, March and April. For example, 20 19-0 1 represents the earliest consumer groups in 1 month 1, February, March and April.
To sum up:
Vertical analysis of users in the same life cycle stage, so as to compare the changes of similar groups with time, as can be seen from the above figure, the user retention rate is decreasing with time.
At the same time, by comparing the same period groups at different times, we can see that the retention rate is high or low. As can be seen from the above figure, it fell in February 2065438 and rose again in April 2065438.
It may be that the user activity led to this result at 20 19-03, thus verifying that the activity improvement has achieved obvious results.