Presented by AWS Machine Learning


About 2.5 billion people around the world are underserved by traditional financial institutions. For traditional banks and loan companies, these individuals technically don’t even exist: According to one World Bank estimate, approximately 68% of adults have no credit data recorded by any of the private bureaus, and therefore no credit score. And in most of these areas, national IDs do not exist. Without basic access to the credit-based economy and a way to prove citizens’ identities, about 85% of the world’s transactions are cash-based, limiting or preventing their access to the global economy.

This has created a frustrating vicious circle. Without financial data, it’s hard for financial services companies to identify this population, hard to reach them, and hard to transact with them.

But financial services company Tala recognized that with machine learning powered by AWS, they could find this population of unbanked consumers and serve them with a modern credit infrastructure built from scratch. They’d be the first to offer loans to this population of previously unbanked consumers, and unlock their financial potential in the marketplace: their ability to participate as both buyers and as business owners, and contribute to their community.

“Their credit worthiness is difficult to ascertain, but they have a tremendous amount of purchasing power and potential to harness,” says Shivani Siroya, CEO & Founder at Tala. “We saw the value that these unbanked people could bring to the formal marketplace and their community if we could unlock their ability to participate, and of course, the immense benefit to themselves and their families.”

Tala’s mission, Siroya says, is to enable financial agency for the emerging middle class, globally. 

To date, Tala’s innovative machine learning solution has found, scored, and offered more than 5 million customers $2 billion in loans, Siroya says.

How Tala works using machine learning

At the core of Tala’s strategy is a key insight: Machine learning is a powerful tool – but only if you’re listening to your customer, and only if you understand the problem.

In Tala’s case, Siroya had the opportunity to learn first-hand about the challenges of the unbanked, and the difference small loans made in their lives and in their communities, over the course of her career at the UN Population Fund and in microfinance in India. Listening to those people and their stories, she actually found herself willing to invest her own money in their futures.

“What unconsciously happened for me is, by working and living in those markets, because I believed in them, I started lending my own capital to them,” explaines Siorya. “And then it occurred to me that the way I was choosing which businesses to invest in or lend to was based on both behavioral data and capacity data. Which led me to this question: How could we find a more seamless, as well as scalable, source of this information?”

The source of that information lay in mobile phones, which are ubiquitous, even in unbanked populations. At the same time, one of the traditional limits of microfinance is the cost of underwriting and making the decision to approve or decline. And this can be costly if you make a small loan amount.

To harness the ability to make decisions based on behavioral and capacity data, and solve the microfinance cost problem at scale, Tala developed an innovative machine learning solution, built on AWS infrastructure. They’ve adopted automated processes end to end, leveraging machine learning at each step of the decision process to approve or decline a loan, to determine loan amounts, or make decisions concerning collections and acquisitions. The entire flow uses machine learning to automate as much as possible and reduce the cost — and therefore serve more customers at a better price.

“In a nutshell, this is the mission of machine learning at Tala,” explains Nabil Abdellaoui, data science manager at Tala. “Every time we improve the quality of a decision in the model, we can serve more customers, and better. It’s a great opportunity for the data science team at Tala, because we see the impact of our work on these algorithms every day.” 

To creatively solve the challenge of no credit history, the algorithm uses an applicant’s smartphone data which provides important eligibility information. This includes device type and ID, the year of the operating system, and the apps on the phone. It also analyzes how a customer interacts with the Tala app – for example, which pages they visit and how they arrive at those pages, or whether they read the terms and conditions. Once a customer has taken a loan, their repayment behavior is the most important factor for future lending decisions.

The algorithms are based on a variety of modeling techniques, Abdellaoui says. For credit decisions, the data scientist teams chose gradient boosted trees, a machine learning technique used for predictive modeling. Fraud defenses generally rely on anomaly detection techniques to identify suspicious behavior or incidents. Collection treatments use uplift modeling — techniques that don’t predict the outcome, but try to predict the best treatment to reach the desired end result.

As the company’s customer base grew both vertically and horizontally, they found that Amazon SageMaker was the natural evolution in graduating from gigabyte data sets to terabytes, since they were already AWS customers, Abdellaoui says.

They use SageMaker today for monitoring, model deployment, and to do automated model retraining to react very fast to changes in the world. For example, the COVID-19 situation required Tala to update algorithms very quickly without wading through a long development process.

The biggest challenge they faced as they built the platform was connecting the dots between data science, R&D, and the prototyping phase with the actual productionizing of machine learning.

“SageMaker helped doing this, because it allows data scientists who don’t have very deep DevOps or data engineering skills to get to the result and get the model deployed in production very robustly using AWS standards,” he explains.

The intersection of technology and human understanding

The foundation of Tala, and their innovative machine learning solution is the balance they struck between leveraging the technology to enable a whole new market and looking behind the data, sitting down to listen closely to their customers, in order to fund their dreams.

Along the way they’ve helped customers like Caroline, an entrepreneur who now runs an online women’s boutique in Nairobi. With help from Tala she grew her loan capacity from $10USD to $300USD.

For Nico, an aspiring K-POP star based in Manila, his ultimate goal was to make it big in Korea, like the most successful K-POP stars. That dream seemed impossible, until he learned about a dance competition coming to his city, on the lookout for breakout stars. He used his first Tala loan to create custom-made professional costumes for his band, so they could enter strong and bring his ambition to life.

“What it always comes back to is the context behind the data,” Siroya says. “It’s about ensuring that machine learning requires a core component of user research and understanding the customer firsthand, and then using the technology to bring those stories to life.”

Dig deeper: See more ways machine learning is being used to tackle today’s biggest social, humanitarian, and environmental challenges. 


Sponsored articles are content produced by a company that is either paying for the post or has a business relationship with VentureBeat, and they’re always clearly marked. Content produced by our editorial team is never influenced by advertisers or sponsors in any way. For more information, contact sales@venturebeat.com.