A group of engineers and credit specialists are working in a converted office building off Shoreditch High Street, where the coffee is served in cups bearing the company’s logo and the whiteboards are covered in decision-tree diagrams that would require a data science degree for a visitor to fully understand. It may seem straightforward, but a mortgage broker with twenty-five years of experience would find it difficult to understand. They are determining whether a person should be granted a loan based on how they now handle their money, in real time, using the behavioral signals that come through their bank account each week, rather than what they have borrowed and repaid in the past.
The idea behind the standard credit check made sense when it was first created, but as digital banking grows more prevalent in how most people manage their finances, it becomes less logical every year. The idea is that a person’s prior debt repayment history is the best indicator of how they would repay future debt. That’s not incorrect; payment history is important. However, it poses a structural issue for a wide range of borrowers who have not had the chance to establish a traditional credit history, such as recent immigrants, young adults who have never had a credit card, and independent contractors whose income patterns appear irregular to a scoring algorithm intended for payroll employment. Regardless of how well they actually handle their money, some 5.8 million persons in the UK have credit files so thin that major lenders usually reject them automatically.
The behavioral AI approach begins with a distinct premise: that a person’s buying habits, cash flow patterns, and regularity of payments contain predictive information about creditworthiness that is at least as reliable as a static credit score, if not more so. Fintech lenders have access to a real-time feed of an applicant’s account behavior through open banking, the UK framework that permits users to share their transaction data with authorized third parties. Do they regularly pay their utility bills? Do they run a negative balance before to payday or keep a positive balance throughout the month?
Over the last six to twelve months, have their recurrent payments—such as rent, subscriptions, and cellphone contracts—been consistent and on time? When machine learning models trained on past applicant outcomes analyze these signals at scale, they generate a risk assessment that can identify risk in applicants whose credit bureau scores appear normal but whose behavioral patterns indicate stress, while also approving borrowers who would be rejected by traditional underwriting.
Given the intricacy of what it is attempting to regulate, the FCA has been observing this development with a mix of sincere interest and regulatory caution. Increasing underprivileged people’ access to financial goods is a policy goal that the regulator takes seriously, demonstrating the potential of behavioral AI in credit. There are actual risks as well. When a black-box model is used to make lending decisions based on transaction data, it may contain biases that are not visible to its operators. These biases may result in systematically unfair outcomes along the dimensions of geography, spending patterns associated with ethnicity, or behavioral signals that disadvantage specific demographic groups without explicitly including demographic characteristics. Businesses using these models must give explainable decision logic in order to comply with the FCA’s sandbox and live-testing cohort programs. This means that if someone is rejected, they must be able to comprehend the reasoning behind the decision in meaningful terms rather than simply being informed that the algorithm rejected them.
The engineering becomes very challenging when it comes to the explainability criterion. The most predictive behavioral AI models are typically the hardest to understand, such as intricate neural networks or gradient boosted trees that find patterns in transaction data that are difficult for a human analyst to decipher. Easy-to-understand models, such as simple logistic regression on a limited number of interpretable variables, are typically less accurate. London-based fintech companies operating in this field are developing two-layer systems: an explainability layer that converts the model’s logic into language that regulators and rejected applicants can comprehend, and a predictive model that produces the risk score. The scholarly literature on explainable AI hasn’t adequately addressed the question of whether those explanations are a credible post-hoc rationalization or truly reflect what the model is actually doing.
Here, the background of market rivalry is crucial. One of the major credit bureau incumbents, Experian, acknowledged the validity of the alternative data argument and the necessity for the incumbents to address it by launching Experian Boost, which enables UK consumers to add utility and subscription payment data to their credit reports. In 2022, Apple stunned the market by acquiring Credit Kudos, a London startup that developed an open banking-based credit evaluation platform. This move implied that behavioral transaction data was valuable enough to draw in businesses with very wide addressable markets. The Shoreditch-based fintech companies developing this technology are competing in a fast-paced market that has drawn interest from organizations with significantly larger balance sheets.
