The Interplay of Experimentation and ML to Aid in Repayment of Micro-Loans in Sub-Saharan Africa
Blogs from ODSC SpeakersConferencesData Science for GoodMachine LearningModelingData Science for GoodmicroloansMLposted by Brianna Schuyler August 16, 2019 Brianna Schuyler
Editor’s Note: Brianna is speaking at ODSC West 2019 and ODSC Europe 2019, see her talk “The Interplay of Experimentation and ML to Aid in Repayment of Micro-Loans in Sub-Saharan Africa” there
Imagine through a twist of fate that rather than living the life that brought you to reading this today you were born in rural Uganda to subsistence farmers. You didn’t have much access to school and now, as an adult, you earn your living through farming cassava and maize. On average, you earn about 60 USD per month, but it’s mostly split across two harvest seasons a year. With this cash, you have to budget for all your household needs: food, charcoal, clothing, school fees, healthcare, phone charging, and oil for a kerosene lamp to work at night.
[Related Article: Watch: Using Data for Good]
The smoke from the lamp causes coughing, and the light isn’t bright enough to see well, so your kids aren’t able to study past sunset. The risk of fire is also high, only last week your neighbor tipped his lantern over, almost burning down his hut. An efficient solar panel and power system would provide cleaner, safer, and brighter light at night, as well as the ability to charge a basic cell phone to keep in touch with family. However, even a small reliable kit costs around 100 USD, and you have no documented financial history to apply for a loan.
This is the situation for around 600 million (!) people in Sub-Saharan Africa. Easy-to-use digital financial technology is allowing companies like Fenix International to offer solar home systems on loan; our customers can pay off their kits with small payments (as low as $0.14 per day) over the course of 1 to 3 years. Even with affordable financing, our lower-income, rural customers can face significant challenges to keep up with their payments. Seasonal income from farming, seasonal expenses like school fees, and shocks like medical bills can make cash flow and payments quite irregular. The customer may experience challenges making payments due to poor availability of mobile money agents or if the mobile money network is down. And if there is an issue with the product itself, such as a faulty solar panel or—as often happens in Uganda—if a rat has chewed through the wires, the customer will resume payment only once the issue is diagnosed and resolved.
All of this uncertainty and complexity makes data an essential part of helping Fenix International customers stay on track with their repayment—through analytics, predictive modeling, and A/B testing around interventions. Data from multiple different sources, including customer demographics, regional information, IoT data from the solar kit itself, as well as the customer’s own past repayment patterns, are used to understand when and how to intervene when one of our customers is falling behind on their loan, and to work with them throughout their repayment process.
One of the ways Fenix interacts with our customers is through our call centers, which service dozens of local languages and can be very resource-constrained during peak hours and as we grow quickly in a market. We use predictive modeling to bring efficiencies to our customer experience by allowing us to identify customers who may need a call from us, and to prioritize calls to customers who are both likely to answer and who are more likely to make a repayment or have an issue resolved after receiving a call from our customer service team.
[Related Article: 4 Examples of Businesses Solving Problems with AI]
Fenix uses data and analytics to identify and solve specific pain points in the customer journey, as well. We found that customers who are late to their first digital payment (after their initial deposit) are more likely to default on their loan. As a result, we created a random forest model to predict who would be late to their first payment and are now testing how much a preemptive call from our call center increases initial payments in that group. These data will also serve to train a second model to predict the customers who are likely to respond well to a call, allowing us to use our exceptional call center representatives effectively.
We are constantly iterating and experimenting with different initiatives and behavioral interventions setting up control groups and measuring effects and ROI. We go into the field and interact with kit-owners to better inform our feature engineering and selection and to evaluate the customer journey. Our IoT pipeline analyses millions of pieces of information sent daily from our devices to understand product usage and if this usage predicts future payment behavior. At ODSC West 2019 and ODSC Europe 2019, I’ll talk more about the mix of experimentation and ML that we use to increase financial inclusion and adoption of sustainable energy sources, as well as some of the issues that we’ve faced and hard lessons that we’ve learned working in this really exciting and unique space.