Background
Our client is a Canadian instant loan provider specializing in non-credit check loans and offering consumers alternative credit. In the loan approval process, the past 3-months of user transaction history is collected and underwriters utilize this data along with a generic rule-based scoring method to make decisions on whether a client should receive a loan.
Problem Statement
The objective of this project was to determine if a more accurate scoring method for non-credit check loan approval could be developed. As the majority of loans are denied, an accurate predictive model would be able to assist underwriters by narrowing down the pool of potential clients to those who are most likely to pay back their loan. Therefore, the Beam Data team aimed to develop a set of models for predicting the capacity and propensity to repay their loan.
Methodology
Two modelling approaches were identified for this project:
- Creation of a custom scoring method using a rules-based system
- A rule-based system is often favoured in scenarios where the data being used is stable and the exact logic is known; however, this comes at the cost of flexibility once deployed as the rules are hard coded
- Creation of a custom scoring method using machine-learning (ML)
- A ML method is favoured in scenarios where the volume of data is high, relatively variable, and the exact logic is not known
In both scenarios, the bulk of the development time will be spent on engineering features centered around one’s finances. These features will then be used as inputs for our models to allow for the differentiation between a ‘good’ and ‘bad’ client.
Customer Journey
The customer journey to loan approval involves several key steps. First, the client creates an account on our client’s platform; a third-party service is then used to retrieve the client’s banking history and with the available financial information, the underwriter will make a decision on loan approval. If the loan approval is successful, terms of the loan – including the loan amount and repayment schedule – are offered to the client; however, in the event where the loan is denied, the client may apply again in the future once they are in a better financial situation. Furthermore, depending on the client’s ability to pay back the full loan amount, the client may be considered for future loan applications.
Conclusion
In the non-credit loan space, differentiation between a ‘good’ and ‘bad’ client is challenging due to a limited banking history time window. To tackle this data problem, the Beam Data team created over 25 customized features centered around a client’s recent finances. With a set of customized features, along with the loan approval and denial history, we were able to develop a rule-based and an ensemble ML model. A customized prediction function was also generated so that only users with a probability of 90% or greater were denied a loan. Future works for this project will involve further feature engineering and optimization by the Beam Data team to achieve greater F1-scores.
