Effective Credit Scoring Using Limited Mobile Phone Data

There has been a recent explosion of companies providing micro-loans through digital media in many developing countries. This explosion is fueled by the need for quick and convenient loans, and enabled by the vast adoption of mobile phones and mobile money. To screen borrowers, these digital lenders typically collect massive amounts of data, such as communication patterns, data on social media activities, and detailed mobile phone usage from their customers. These data present a number of potential privacy risks to borrowers. In this study, we demonstrate that accurate credit-scoring models can be trained using only airtime recharge data, which we argue is less invasive to the borrower's privacy than the typical model employed by lenders. We tested this approach through a partnership with an airtime lender in Africa that made it possible to run a side-by-side comparison of an airtime-only model against a model that also incorporated past loan data, as well as the current model used by the lender. In several tests, our model, which used limited data, performed at least as well as alternative models. These results suggest new opportunities for digital lenders to build reliable credit scoring models that reduce the privacy risks posed to their borrowers.