Entity registration (for linkage of existing loan/user identifier with collected anonymous data) and prediction requests are performed via REST API calls. In addition, Clair also features Clair Dashboard which is the part of Clair deployment. Its usage is optional and depends on the customer needs – all scoring possibilities are fully accessible via API.
Overall, Clair architecture and communications can be summarized at the scheme below:
ThreatMark Clair represents the modular scoring solution for online consumer lending businesses. By coupling deep
In contrast to traditional credit risk models, which operate mainly with historical user data and recommendations provided from credit bureaus, ThreatMark Clair
Overall, differences between traditional scoring and scoring using Clair on different stage of the loan funnel can be summarized on the scheme below:
From collected user interactions:
Clair extracts hundreds of features, each partially describing user behavior. Combined, these features result in multidimensional vector that represents digital fingerprint of each user’s online behavior. Such fingerprints then can be compared with similar fingerprints form other users.
Supported by additional information about final loan state (e.g., paid back or fraud), selected behavior fingerprints are combined to form several target groups of interest. Using standard classification approach, Clair trains its models to find correlation between user behavior and possible outcomes (e.g. target group – fraud vs control group – paid loan). Then Clair will assign a score to each new user, indicating how likely (s)he belongs to the target group, for example, how likely (s)he turns out to be fraudsters or insolvent.
Such a prediction is possible due to the minor differences in the online behavior of ordinary people, who seriously intent to pay the loan back, and behavior of fraudsters, who use fake identities or intentionally don’t expect to return the loan.
As an example of such verifications can be:
The final scoring system can be represented by several custom
During prediction every Clair model issue a probability score between 0 and 1. This score can be interpreted as a measure of how likely is current user belongs to the target group of interest (e.g. fraudster, insolvent, suspicious email, etc) according to his/her online behavior or related digital attributes. I.e. Clair is trying to describe the user as a human being, rather than
By default, Clair is trying to adjust score threshold for optimal precision/recall ratio to achieve as less false positives as possible. Such policy results in actionable score upon which customer can immediately react. However, the most customizable scenario would estimate inclusion of every Clair model score as additional variable into already existing risk engine.