Risk control process
Four verifications: name, ID card, bank card, mobile phone number
Live detection: face recognition
Application strategy: Exclude abnormal groups
Application score card + machine learning: grade division
Quota model: Evaluate quotas of different groups and pricing
cash withdrawal strategies : Verify whether it is my cash withdrawal
management during the loan: detect the credit change by detecting key indicators, and may require early repayment.
Post-loan collection: early text message, later call collection
There is no data for the project cold start
operations early
- Strategy comes from experience judgment (setting rules)
- Blank model of
mid-term business - Strategy comes from data analysis
- Model construction period
Late business - Strategy complements the model
- Model maturity period
Other strategy methods:
strategy analysis (decision tree)
strategy tightening: decision tree method reduces overdue rate (preferably overfitting)
strategy relaxation: edge trend analysis method
in the score card, divided into A, B, C card , Separately recorded the different links of credit evaluation, corresponding to before loan, during loan and after loan.
(Important) When doing data analysis, you first need to determine or transfer the research questions to measurable variables, and in the process of analyzing the variables, you need to control the influence of other external factors
Scorecard Development-Decision
1. Binning——Purpose: To eliminate internal disturbances.
Continuous variable binning method: equal distance, equal depth, decision tree (minimum box number is greater than 5% of total number)
categorical variable binning method
2. Calculate the IV (information value)
greater than 0.02, and
then filter the variables again (the score card is generally 15 variables and the following is more reasonable).
Variable clustering: select (1-R ^ 2) the smallest group variable (eliminate collinearity) and
then forest: importation vs. variable Sort
3. Calculate WOE value
5. Model evaluation-ROC (common evaluation method: around 0.76)
Model evaluation stability-PSI
