- Using Microcred’s loan application data and internal loan status information, build statistical models to help predict customer default and better inform decision making about lending to make it more efficient and inclusive.
- Identify whether loan application data and factors such as past repayment behavior or late payments during the early stages of a loan cycle could predict default.
- Demonstrated that Microcred’s loan application data could be used to help predict default and identified repayment behavior as a significant predictor for default.
- Created predictive models and an analysis pipeline to build default scores for customers at different points in a loan cycle, helping make the loan application process more efficient and effective in determining default risk.
- Built a Monte Carlo simulator to help ascertain expected revenues and expenses associated with customer lending. The financial model could potentially be used to inform the cost-effectiveness of interventions like mitigating defaults and improving customer retention.
French digital finance group Microcred works to contribute to the growth of local economies in developing nations by offering simple accessible financial services. The organization provides lending to nearly a half-million micro-entrepreneurs in eight African countries and China that either lack or have weak guarantees, and would otherwise be unable to access financial markets.
Lending to small entrepreneurs in most developing countries is an expensive and high-risk endeavor. Without credit bureaus to screen borrowers for creditworthiness, lenders must invest their own resources to assess each loan applicant’s ability to repay. A costly, time and resource-intensive process, Microcred currently conducts field visits with potential borrowers individually to evaluate their financial situation and collect necessary information like income, sales, expenses, etc. The collected loan application data, along with existing internal data Microcred has on their customers’ repayment behavior, is used to determine potential risk for default and decide which customers qualify for loans. However, Microcred does not apply a precise formula to the customer data to determine acceptance or rejection of a loan application; it is decided by a committee in a more ad-hoc manner.
Microcred sought DataKind’s help to create a data-driven approach to lending, and also investigate factors related to loan default. They wanted to explore the value of the data they collected from customers to date and build a predictive model that could be used to help determine default risk. Furthermore, Microcred wondered whether or not things like repayment behavior could serve as indicators for default. The goal of the engagement was to provide tools and recommendations that would help Microcred streamline its lending process to make it more efficient and more inclusive.
In a DataKind project sponsored by IBM, a DataCorps team of volunteers including James Beveridge, John Murray, Masha Westerlund and the project’s Data Ambassador Raluca Dragusanu, were tasked with using Microcred’s current customer data to develop credit scoring models to predict default at different stages in the loan process and answer a set of questions that could better inform the loan application, data collection and approval processes.
The DataCorps team focused on Microcred’s lending program in Senegal where Microcred started operating its first branch in Dakar in 2007. In Senegal, Microcred had a portfolio of more than 212,000 customers amounting to a loan portfolio of 90 million euros, and employed more than 600 people. The data sample for this project included the years from 2008 to 2015 and 109,265 loans.
The team combined predictive modeling, along with with financial analysis and simulations, to identify variables from customer loan applications and repayment history that are predictive of default, and to calculate customer lifetime value. It was found that using both loan application and customer behavior data to identify customer financial variables maximized the predictive power for default. How timely a customer was in making repayments to its previous loan from Microcred proved to be a significant predictor for default in the current loan. The team also discovered that lateness in making the first repayment to Microcred was a meaningful early warning sign for default. Microcred can use this information to identify customers experiencing financial difficulty and intervene to provide restructured payment terms to prevent default altogether.
A Monte Carlo financial simulations of customer lifetime value was also created to map customer characteristics into expected profitability over multiple loan cycles. With this tool, Microcred will be able to make lending decisions informed not only by expected default (obtained from the credit scoring model) but by the costs of servicing a customer and the potential for customer retention for multiple loan cycles. The Monte Carlo financial model could also be used to run a “what if” analyses to study the cost-effectiveness of various interventions, to help mitigate defaults and improve customer retention.
The models and findings from the project have the potential to help Microcred to advance its mission of offering simple, accessible financial products and services to people who would not otherwise have access to the financial sector. By adopting a data-driven approach to lending, Microcred will be able to reduce default and streamline their loan application process, allowing them to allocate more resources to expand financial access to customers that are currently too costly to serve and who are otherwise excluded from the financial system. Microcred will also be able to improve loan terms and offer lower interest rates for both existing and new customers.
However, because Microcred currently only keeps data for accepted applications, the modelscan only be applied to customers already being granted a loan and cannot be used to determine whether or not a loan should be granted. As such, the team recommended that Microcred store all applicant information moving forward, regardless of application outcome. Doing so would allow Microcred to create credit scoring models that are free from sampling bias.
The DataCorps team also provided recommendations on additional ways Microcred can leverage its current data sources more effectively and identified specific aspects of the lending process that would benefit greatly from data-driven models. Suggestions on collecting future data and ideas for further data analysis and exploration of additional variables were also provided.