Importance of machine learning for Numida
Numida is a financial technology start-up in Uganda providing unsecured credit to small business owners. Users enter their sales and expenses into the app in order to apply for unsecured loans. Numida’s hypothesis is that a user’s app usage can predict the probability of loan repayment. As such, users must use the app for at least 3 days before they can receive a loan in order to provide Numida with sufficient data to gauge a user’s credit worthiness. Numida’s use of data to make predictions is at the core of its product offering and is one of the most common applications of machine learning.
Numida’s activities in machine learning
Accuracy of predictive models depends on data – there needs to be a large enough sample size and variation in the data for machine learning to be useful. Numida thus needs to a large number of users who have different usage patterns on the app (e.g. frequency of data input, type of data entered, whether or not they apply for a loan, and the loan size they take out). There also needs to be a variation in outcomes, i.e. some users will need to repay the loan and some will need to default, in order for machine learning to be useful in predicting repayment probabilities.
For the past year, Numida has been struggling with getting people to use the app and apply for loans. This means that the sample size has not been large enough for Numida to use statistics to identify app usage behaviors that are correlated with higher probabilities of repayment, since app usage has been low and the number of borrowers even lower. As such, management has focused on making the app more user-friendly to increase user retention and experimenting with the loan eligibility criteria to broaden the borrower base. Only with more users and borrowers will Numida be able to use machine learning for predictive analysis.
Given that it is an early stage start up that has yet to find product-market fit, Numida is more focused on short term strategy and does not yet have a 2+ year strategy.
The biggest challenge Numida faced as of August 2018 was user retention and getting loan applicants, an issue ill-suited for machine learning given the low user base (<150 weekly active users). With this small sample size, a simple before vs. after analysis is sufficient to determine whether a change in the user interface or the loan product resulted in any uptake. Some analysis has been conducted in this vein and Numida needs to continue to build a culture of monitoring the effectiveness of their client acquisition strategies.
Once Numida finds product-market fit and acquires a steady stream of users and borrowers, it can use machine learning to streamline its loan application process. Currently Numida relies on humans to check each loan application submitted on the app before transferring money to the user. As part of the due diligence, sometimes loan officers need to contact users to request them to reupload documents because of poor photo quality, causing delays in loan disbursement. This friction in the customer journey causes users to wonder during the delay if Numida’s value proposition of unsecured credit is a scam and discourages users from persevering in building up a habit of entering sales and expenses into the app. An ability to leverage machine learning to conduct basic due diligence of photo quality when users upload photos of their documents and provide immediate feedback on whether it needs to be retaken could streamline the customer journey and result in higher user retention. However, this will likely be worth the investment once Numida has sufficient application volumes.
There are several open questions regarding the applicability of machine learning within the developing world. Ugandan’s digital footprint is still sparse compared to those residing in the developed world due to Uganda’s cash-based economy and the careful rationing of internet use due to the cost of internet data. Numida’s hypothesis of using data to predict a user’s creditworthiness is reliant on inputs that are hard to collect – users are reluctant to build a habit of inputting their sales and expenses into the app, yet the opportunity to collect user data is limited since internet use is still fairly limited. In such a situation, how can Numida get vast amounts of data cheaply and in a more convenient way for users that would enable Numida to properly predict probabilities for loan repayment? (787)
 Numida. (2018). About. [online] Available at: http://www.numida.co/about [Accessed 13 Nov. 2018].
 Yeomans, M. (2015). “What every manager should know about machine learning.” Harvard Business
Review Digital Articles, 7 Jul. 2015.
 Higginbotham, S. (2018). http://fortune.com. [online] Fortune. Available at: http://fortune.com/2015/06/15/facebook-ai-moments/ [Accessed 13 Nov. 2018].