1. Project Overview and Purpose

In the cash loan business, there are two main risks: credit risk and fraud risk.

Credit risk mainly evaluates the borrower's repayment ability and willingness to repay, thereby determining the loan amount (it can also determine the loan interest rate),

And anti-fraud What we are directly facing are fraudulent users. This type of user has no other purpose than to defraud money and refuse to pay it back (including wanting to pay it back at first but not paying it back later).

If you directly use rules to fight fraud, there are three limitations:

1. The strategy is relatively strong, and the hit is directly rejected, and the deep relationship between strategies cannot be considered;

2. It is impossible to give the user's fraud risk;

3. The transfer of users from credit risk to fraud risk is not considered.

The anti-fraud model is designed to use machine learning to improve the current shortcomings of simply relying on rules to reject people.

2. Implementation plan:

There are two types of fraud, one is direct fraud, and the other is the transfer of credit risk to fraud risk (the user originally had a weak willingness to repay, but then As time goes by, the weak willingness to repay disappears

There is also a situation where there is a willingness to repay but no ability to repay, so when selecting features and tags, consider these directions< /p>

Types of anti-fraud scorecards:

1: Small amount

Before loan (mainly to obtain characteristic training before loan)

During the loan (mainly adding data suspected of fraud generated during the life cycle of our platform)

2: Large amount

Before the loan (mainly based on what can be obtained before the loan) Feature training)

Daizhong (mainly adding fraud-suspected data generated during the life cycle of our platform)

1. The anti-fraud score is set to 350-970 for each user.

a). If the score is less than 300, it will be directly considered as a fraudulent user, and it will be directly rejected online

b). If the score is between 500 and 700, there is suspicion of fraud, but we are not sure. It is not a complete fraud. A small part of these users will enter manual review, and most of them will enter other anti-fraud models

c). Users with 700 points will be admitted directly, and then enter other strategies and credit models

2. Feature selection is divided into the following parts:

1: Gang-related fraud features

a). Combined with device-related features (such as how many mobile phone numbers are associated with the device , how many ID cards are associated with the device, how many devices are associated with the mobile phone, etc.)

b). IP behavior related (whether the IP is an abnormal IP, the number of login IPs)

c). Gang-related (such as how many prospects are in the gang, how many prospects are in the primary contacts, how many prospects are in the secondary contacts, the number of blacklists in the primary contacts, the number of secondary contacts in the blacklist, whether there are intermediaries in the gang)

2: Personal-related fraud characteristics

1. Detailed bill data (such as the number of calls at 0 o'clock, the number of calls with direct contacts in a month, etc.)

2. Risk factors in the anti-fraud strategy in the decision-making flow (such as hitting the court execution list, hitting the three-party blacklist, etc.)

3. Three-party fraud data (such as Tongdun risk rating)

4. Behavioral data (this part only applies to loan anti-fraud, user expected times, etc.)

5. Data generated by anomaly detection

3. Label Choice:

a). Manually mark the user after the loan

b). Manually mark the user before the loan

c). The policy directly rejects the user due to anti-fraud< /p>

d). The first order of installment products is overdue for more than 14 days (first order, first issue)

e). The maximum number of days for installment and single-issue products to be overdue is 30 days

f). Blacklist users

4. Data fusion

If the data is not easy to fuse, consider making the data that is not easy to fuse into a sub-model. The model scores are brought into the main model as main model features