How AI could transform financial services in emerging markets

Financial inclusion is a global challenge. Solving it has enormous benefits, not only for the financial institutions that can offer new customers access to products and credit they may previously not have had, but for empowering the unbanked, enabling them to access financial services, and assisting them in financial planning while offering them hitherto inaccessible means to pursue their aspirations.

A society with greater financial inclusion is also good for the wider economy. McKinsey estimates that using digital financial services for inclusion alone could add US$3.7-trillion to an emerging economy’s GDP in less than a decade.

The World Bank says that an inclusive financial system is fundamental when it comes to reducing extreme poverty, promoting sustainable inclusive economic growth and development, and boosting shared economic prosperity. Why? Because greater financial inclusivity allows those previously excluded to borrow, save, invest and start businesses, while also positioning themselves to weather socioeconomic disruption.

More than two billion people globally do not use formal financial services, in part because financial services providers haven’t found the right way to approach them, or been sufficiently motivated to. This is because the unbanked are deemed expensive to provide services to, and it’s assumed there is little prospect of them eventually providing the sort of upside to make them worth the effort.

ResearchGate found that despite improvements to African banking services in recent years, the continent’s ratio of private credit to GDP remains far behind that of other markets. It averaged 24% of GDP in sub-Saharan Africa in 2010 and 39% in North Africa, which is far lower than the 77% average of all other developing economies, and the 172% average of developed ones.

The problem is no better if one zooms in on individual access to finance. A paltry 23% of adults in Africa have a bank account. But it’s worth noting this figure varies considerably between countries and regions. For instance, in South Africa, the figure is 51%, while it plummets to 5% in the Central African Republic.

Fewer than a third of adults globally have access to any sort of credit bureau, even though the World Bank estimates that two-thirds of the global unbanked nonetheless have access to a mobile device. This is key because lack of credit history isn’t only a problem for adults; it’s one millennials looking to buy cars or houses, or fund education, also face. But new approaches can help alleviate this problem. For instance, machine learning and big data can be used by new fintech players (or established ones) to create credit scores for consumers who don’t have sufficient data points in the traditional financial system.

Alternative data

Generally speaking, to secure a line of credit, a person needs a financial history, and a detailed one at that — simply opening a current or savings account is seldom sufficient. In those places where children or teenagers get accounts opened for them by their parents, and where those same parents are willing and able to act as guarantors, access to credit is taken for granted. But in cases where the parents themselves have no conventional banking products, credit is all but impossible to come by.

Artificial intelligence can change this by allowing financial services providers to make use of alternative data sources to assess creditworthiness. For individuals, meaningful inferences can be made by looking at data on their smartphone (with their permission) which can reveal financial behaviour, from call activity to app usage, and even which applications a consumer uses most. In agriculture, meanwhile, satellite images can be used to estimate past and future income from farming and make decisions about loans accordingly.

This opens the door to new avenues of lending: for example, smartphone-based microlending, bereft of the usually prohibitive and punitive interest rates such lending models tend to use to insulate themselves against risk. Because AI can construct pseudo networks of similar people, and combine complex network analysis with representation learning, it can also make assessments with ethical concerns and privacy regulations in mind, while avoiding the sort of discrimination that often dogs human-based decision making.

The ever-increasing affordability and adoption of smartphones opens up entirely new market segments to financial services providers

For instance, the time and location of transactions — and other spatiotemporal traits — have a surprisingly wide range of predictive value. Similarly, the social networks (or social graphs) of individuals can provide impressively accurate and actionable signals of credit quality.

The ever-increasing affordability and adoption of smartphones opens up entirely new market segments to financial services providers. The increasing use of social media provides another avenue for alternative data because the connections between people can reveal all sorts of probabilities. For example, a person’s most contacted friends can influence the likelihood of them abandoning a service if those friends have previously done so themselves.

Alternative algorithms

In conjunction with alternative data, alternative algorithms can improve risk assessment when dealing with consumers with limited financial data to offer. Logistic regression is already commonplace in banking, but by adding additional vectors like random forests, assessment models can account for features previous models might discount or miss and be better equipped to identify those that are statistically relevant.

For example, a California-based start-up called Tala uses 250 data points, including consumers’ online behaviour, cellphone and data usage, and other metrics, to assess their desirability as clients. The process begins with identity verification: is the prospective customer who they say they are? Thereafter, it uses supplied data to make a decision on a customer’s loan request.

Tala customers who are approved for loans are incentivised to pay them off in less than the allotted period, something microlenders often penalise customers for, and personal data is deleted after the loan is repaid. Instead of getting users to resubmit the data they did originally for any subsequent loans, their previous repayment behaviour is used for assessment.

Similarly, a US service called Wemimo uses utility bill and rent payments to help create credit scores for the estimated 45 million Americans without a traditional credit history, the lack of which generally excludes US citizens from getting loans for cars, houses, or other high-cost purchases or expenses.

This also opens the door for lenders to enter into partnerships with utility providers, service providers, wholesalers, telecommunications providers, or even governments to identify promising data sources for assessing potential customers’ creditworthiness (with those customers’ explicit consent). Alternatively, lenders can help these repositories of user data to become lenders themselves, something many retailers already do by using a loyalty card first to assess spending behaviour and then offering an adjacent credit service to those customers it identifies as suitable candidates.

Another player in this emerging space is Branch As, a credit product that uses mobile payment system M-Pesa as its delivery mechanism. Like Tala, Branch As requires prospective customers to access data on their smartphones via a mobile application they download. AI makes automatic decisions and adjusts interest rates (between 2% and 14% monthly) accordingly. Newcomers to the service receive smaller loans with higher fees, but subsequent loans adjust depending on users’ repayment behaviour, and all successful loans are deposited directly into the users’ M-Pesa account, so they can be available in seconds.

Open APIs (application programming interfaces) from telecoms services providers have made these sorts of credit assessments easier and more reliable by providing third parties with phone usage and bill payment information. In Nigeria, for example, using telco data provides access to credit to whole swathes of the population to whom it would otherwise be out of reach. While fewer than 3% of Nigerians are eligible for conventional bank loans, nearly 50% of them have mobile phones, suggesting a credit scoring system predicated on mobile data isn’t just feasible, it has the power to be massively empowering.

One of the alternative algorithms that make this sort of assessment reliable and useful is reinforcement learning. By using new information to adjust thresholds, reinforcement learning can not only make more accurate decisions on creditworthiness but can also proactively guide customers to optimal behaviour so as to improve their odds of a successful application. Equally, it can be used to flag, preempt and respond to consumers’ financial distress or other risks.

Studies by M Herasymovych et al (November 2019) found that dynamic reinforcement learning systems — that is, those which constantly adapt their credit-decisioning thresholds in response to live data feedback from customers — tend to reduce the sorts of biased results that can lead to eligible customers being misclassified and denied credit erroneously.

Handling imbalanced datasets

A scarcity of data or credit history in some population groups might lead to datasets being imbalanced. If one class is smaller than another, for example, the larger of the two may be favoured unduly. If a model is weighted to non-defaulters, for instance, defaulters may be unreasonably biased against. Imbalance datasets make assessment more difficult and bias more likely to creep in.

Various techniques have emerged to help to mitigate this problem. These techniques range from random-under-sampling and random over-sampling to Synthetic Minority Oversampling Technique (Smote) that tries to provide more accurate filler data by randomly creating data points that are near but not identical to the source ones.  Another solution is changing the performance metrics. These can include using a confusion matrix, adjusting for precision or recall, or using a variety of algorithms and comparing their results rather than relying on a single one.

But there aren’t the only means of injecting fairness into machine-learning (ML) models. Researchers at MIT have used a new technique that reduces bias in image-recognition systems, even when ML models are trained on unbalanced datasets. The solution, called Partial Attribute Decorrelation (Parade) trains the model to evaluate certain metrics separately.

Parade can be applied to other metrics and can similarly flag sensitive ones so that they don’t become reductive measures by which images, creditworthiness or risk are assessed.

Digitising access

Digitising access and creditworthiness, especially in the micro-finance sector, can drive financial inclusion significantly. Aside from reducing bias, it can also improve access, because while many people may struggle to physically get to a branch, most have access to digital channels and platforms.

The solution may lie in a blended or hybrid approach, where some decisioning and loan functions are driven by AI, but others still employ human judgment and interactions. This is particularly relevant where access to digital channels doesn’t translate into comfort using them, and may mean the difference between achieving a competitive advantage and squandering one.

Another company using phone-based assessment is CredoLab, a Singapore-based fintech company which does scoring on lenders’ behalf using smartphone metadata, developing scorecards it can apply to similar applications. CredoLab’s solution has resulted in a 20% higher rate of customer approvals of new clients, a 15% reduction in non-performing loans.

NLP applications in finance

Natural language processing (NLP) is a technology that allows humans and machines to communicate using the sort of language encountered in everyday life. It shows up in everything from smarthome assistants to chatbots and even interactive voice recordings. But it’s also finding a growing number of use cases in finance.

NLP can also be used to assess not only creditworthiness, but trustworthiness, something that’s traditionally hard for humans to do. It can do this by executing deeper analysis on interactions, and by understanding hidden patterns in speech or language that may be too subtle for humans to detect, but which can be recognisable from huge sample sizes gleaned from previous, similar interactions. This approach has been successfully deployed by Microbnk and Capital Float, among others.

In addition to risk assessment, NLP can be used in chatbot advisors that can offer customers credit advice and coaching, or which can assist customers to understand products or specific details about them. NLP also makes it possible to provide services in multiple languages, which in turn can improve consumers’ comprehension and comfort — something that’s especially valuable where complex financial services are involved, like home loans.

Bank of America’s Erica chatbot, for example, is available via the company’s mobile application to more than 25 million customers. Erica acts not only as a financial assistant — reminding customers to make credit card payments or pay bills — but also acts as a financial advisor in conjunction with another product called Life Plan, which focuses on helping BoA customers achieve their financial goals, from purchasing a home to paying for their children’s education or minimising the impact of their existing debt.

AI can do far more than improve financial inclusions. Once consumers are within the financial system, it can help them make the most of the products and services on offer to them. At the same time, it can help financial services providers to reach new customers, provide them with the right products and services, all while reducing risk.

• The author, Dr Mark Nasila, is chief analytics officer in First National Bank’s chief risk office