AI Case Study
Branch grants loans to app users based on alternative data including contacts, social media, and call history and machine learning to assess credit risk
Branch is an Android mobile phone app that uses the data stored on a user's phone to assess their creditworthiness and then deposits the approved loan automatically, allowing returning users with a built-up credit history to apply for larger loans at better rates.
"Branch’s mobile-based application process consists of three steps... First, the loan applicant downloads the application from the Google Play store and grants Branch permission to access his/her handset details, SMS and call logs, social network data, GPS data and contact lists. As part of this step, the applicant then provides his/her country of residence, name, national ID, date of birth and mobile money account details. Second, eligible loan offerings (with transparent amounts and interest rates) are displayed, and the applicant can choose which loan to request. Lastly, the approved borrower instantly receives the cash deposit in his/her linked mobile money account. The entire application and dispersal process can take as little as 10 seconds. Leveraging the smartphone allows Branch to detect subtle patterns of behaviour that correlate with repayment or default. New borrowers begin at the “bottom of the ladder”, receiving smaller loan offers with higher fees."
Specifics undisclosed, but in terms of customer uptake: "Branch has ~250,000 customers, with over US$30m in originated loans and a default rate of ~7%".
"Once the data has been scanned, a machine-learning algorithm automatically and immediately calculates the creditworthiness of the applicant. Smartphone data from new customers is compared with that of previous borrowers to assess the probability of repayment and expected lifetime value, and existing borrowers are rescored each time they apply for a new loan. The machine-learning algorithm continuously learns to improve its ability to assess risk. The machine-learning model constantly learns how to better predict risk as more customers move through the system, contributing to lower default rates with time. It takes ~6 months on average for the algorithm to reach a stable point. The risk assessment model is highly scalable, as the same base machine-learning algorithm can be used in new markets. It will simply learn the appropriate indicators for risk in that new market with time, through exposure to data. It is important to note that while the risk assessment model is highly scalable, the smartphone data that is required for the model is not."
"In Tanzania, credit bureaus only cover 6% of the population
(World Bank, 2016), which leaves 94% of the population excluded from access to formal credit. Branch is a mobile app that works to fill in the gap by using alternative data to assess credit. Branch's first product is in credit, with delivery being through a mobile-based application in partnership with Vodacom M-Pesa. The credit product is an Android-based mobile application that asks users for permission to access and analyse stored data on their phones to credit-score them. As users build their credit history with Branch and positive repayment behaviour is observed, they can “move up the ladder” and unlock higher loan amounts, at better terms. Branch is currently operating in Kenya (as of March 2015), Tanzania (as of April 2016) and Nigeria (as of March 2017), with plans to rapidly expand into new markets. In total, Branch has ~250,000 customers, with over US$30m in originated loans and a default rate of ~7%."
Phone make and model, call, text and data usage logs, contents of text messages, phone contacts, location and movement patterns, contents of phone storage, age and social media data.
"Branch uses over 2,000 data points to make the credit-scoring decision. Answers to questions like the following are used
as key indicators for creditworthiness: 'What portion of an applicant’s phone contacts is stored with first and last names?', 'Does the applicant have outstanding credit at other financial institutions?', 'Are there fraudulent individuals in the applicant’s Facebook network?'".