Fintech startups looking to disrupt payday lending are using artificial intelligence to make loans with rates as low as 6% and with default rates of 7% or less.
AI can make a difference on several fronts, the startups say. It can process enormous amounts of data that traditional analytics programs can’t handle, including data scraped constantly off the borrower's phone. It can find patterns of creditworthiness or lack thereof on its own, without having to be told of every clue and correlation, startups like Branch.co say. And the cost savings of eliminating the need for loan officers lets these companies make the loans at a profit.
Urgency outweighs privacy
MyBucks is a little-known, oddly named Luxembourg-based fintech company that started lending in South Africa but is spreading around the globe.
It’s also doing several things many U.S. banks would like to do, such as identity proofing and enrolling new customers in its lending service through a mobile device and sending loan funds to that device within 15 minutes.
It’s making loans to previously unbanked people with no credit score at rates of 20% for loans of less than six months and 25% to 40% for long-term installment loans. And it’s profitable.
The power behind the lending operation is a credit-scoring engine called Jessie. Jessie analyzes cell phone bill payment history, bank account history (if the person has a bank account), utility bills, geolocation, and credit scores.
“We've built a fraud engine that allows us to credit score quite efficiently, and check whether or not there is any fraudulent behavior,” said Tim Nuy, deputy CEO.
Some of this information, including transaction histories and geolocation, the system pulls from the customer’s own device, with consent.
“Android has no privacy restrictions whatsoever,” Nuy said. “iPhone is slightly less.”
People who are underbanked tend to be unconcerned about privacy. They're more worried about meeting an urgent need for cash.
The software has allowed MyBucks, which has deposit and lending licenses in several countries, to reduce the timeline for getting credit from at least a week to 15 minutes.
“That's transformational,” Nuy said. “That's why we are winning client access and cost even though we're continuously fighting to break the paradigm of people thinking they have to go to a branch.”
Because people don't realize they can use their mobile phone as a bank, MyBucks typically has five or six kiosk-size branches in a market where agents with tablets assist people with the initial application. They teach customers how to serve themselves from a mobile device from that point on.
The cell phone companies MyBucks works with help with the quick identity proofing. In some countries, consumers have to provide a passport to obtain a SIM card. Phone providers and banks won't hand out personal information, but they will confirm basic identity data points.
MyBucks' current loan book is $80 million. The loans range from $5 to $5,000; the average is $250. The smallest loans are short term, up to six months. The larger, longer term loans are installment loans backed by payroll collection mechanisms. They’re used mostly for home improvement, small business, and education.
“Schools in Africa don't generally offer installment-based payments, so people would rather take a loan and pay if off over the year,” Nuy said.
The company has been at a 7% default rate for the past four years, by design.
“The great thing about data science is, we can tell the system what our tolerated risk level is, then the system will tell us which clients to approve and which not,” Nuy said. “And it sets the return rate based on the risk to make sure we get to that default level.”
AI lets MyBucks pull in data components from a diverse set of information points it otherwise wouldn't be able to process, including mobile money payments, income data and utility bills.
“The power of artificial intelligence versus business intelligence is BI is purely retrospective, whereas AI looks forward into the future and predicts — what will this person do based on similarity with other customers?”
AI also helps with an operational reality: MyBucks needs to collect its installment-loan payments from customers in the window between the time their paycheck hits their bank account and when they go to the ATM to withdraw. So it becomes very important to predict someone's effective payday. If payday falls on a Saturday, some companies will pay the Friday before, others will pay the following Monday.
“That's very hard to predict,” Nuy said. “And you have to take into account the different banks — some banks clear in the morning, other banks clear in the afternoon, some banks process same day. …So something very simple, just striking the bank account on the right day and time, makes a massive difference in your collections.”
Leave it to the machines
A branchless digital bank based in San Francisco, ironically named Branch.co, takes a similar approach to MyBucks. It provides its customers with an Android app that scrapes their phones for as much data as it can gather with permission, including text messages, call history, call log and GPS data.
“An algorithm can learn a lot about a person's financial life, just by looking at the contents of their phone,” said Matt Flannery, CEO of Branch, at the LendIt conference Monday.
The data is stored on Amazon’s cloud. Branch.co encrypts it and runs machine learning algorithms against it to decide who gets access to loans. The loans, which range from $2.50 to $500, are made in about 10 seconds. The default rate is 7%.
The model gets more accurate over time, Flannery said. The more information the machine learning system receives, the better it gets at learning from all the patterns it looks at.
“It is kind of a black box, even to us, because we're not necessarily able to understand why it's choosing and who it's choosing, but we know it's getting better and better over time based on a lot of complicated multidimensional relationships,” Flannery said.
Branch.co currently operates in Sub-Saharan Africa and is eyeing global expansion.
In the U.S., however, Flannery noted that the company would be required to provide a single flowchart or explanation for each loan decision.
“That prevents us from making more intelligent decisions and potentially helping people who would otherwise be left out,” Flannery said. “I'm a big fan of allowing innovation in lending, unlike what we do in the U.S."
Flannery said machine learning engines are less discriminatory than people.
“Humans tend to do things like redlining, which is completely ignoring an entire class,” he said. “Machine learning algorithms do [lending] in a multidimensional, ‘rational’ way.”
The company has even considered not including gender as a criterion.
“We're wrestling with these questions,” Flannery said. “I would love there to be a panel or studies done about ways for the industry to self-regulate as this becomes popular around the world.”
Branch.co plans to take AI a step further and use deep learning. “Typically machine learning can be a hands-on process, you have to classify a lot of data and think of new ideas and feature ideas and data sets to classify it,” Flannery said. “But if you just leave it to the deep learning methodology, the classification could be done by machines themselves, which leads to better results in credit over time.”
The black box issue Flannery mentioned has become an issue in the U.S. Regulators have said loan decisions can’t be made blindly — machine learning models have to be able to generate clear reason codes for any loan application that’s declined.
This is why machine learning has been largely irrelevant to lending to date, said ZestFinance CEO Douglas Merrill, who was formerly CIO of Google.
"Machine learning engines are black boxes, and you can't use a black box to make a credit decision in the U.S. or in many other countries, because you can't explain why it did what it did," said Merrill.
ZestFinance has worked with several banks, auto finance companies and other large lenders to create model explainability technology that basically reverse-engineers the decisions lenders’ models make. The software produces a report for adverse action. It will also analyze the model for signs of disparate impact or unintended bias.
"We can open up the model, look inside it, and tell you what the most important variables are and how they relate to each other," Merrill said. "We can call out things like, this variable seems to have a blind spot."
Editor at Large Penny Crosman welcomes feedback at email@example.com.