Lebanon’s banking crisis did something unusual to credit data: it made almost everyone a thin-file customer. Years of frozen deposits, a cash economy and salaries paid through wallets and transfer companies left the formal credit history of most residents frozen in 2019. As lending restarts, through banks, licensed fintechs and microfinance, the question is how to assess a borrower whose last bank loan predates the crisis and whose real financial life since then ran through a wallet.
This guide covers why bureau scores fall short, which data from wallets, salary payments and transactions can rebuild a picture of repayment capacity under Banque du Liban’s new electronic payment framework, how to build explainable models, and how to turn a better score into a product.
Why bureau scores are not enough
A credit bureau score summarises a person’s history of formal credit. It works well for customers who have had cards and loans for years and poorly for everyone else: young adults entering the workforce, newly arrived expatriates, gig and informal workers, small business owners whose finances run through personal accounts, and customers who have simply never borrowed. In Gulf and Levant markets those groups are a large share of the population and a larger share of the growth, and a lender that declines them all is declining its future book.
The thin-file problem is not solved by lowering the cut-off. It is solved by bringing in data that actually predicts repayment for people who have no repayment history, and by building models that use it responsibly.
Alternative data that is usable and lawful
Transaction data from the customer’s own accounts is the richest source: income regularity, spending patterns, balance behaviour, bill payment and the ratio of inflows to outflows predict repayment better than most demographic variables. With open banking consent, a lender can see this across the customer’s banks, not only its own. Telecom data, where the regulator and the customer permit it, adds tenure and payment behaviour on a bill almost everyone has. Utility and rent payments, salary data through government wage systems, e-commerce and wallet history for merchants, and, for businesses, invoicing and point-of-sale data all contribute.
Two rules keep this lawful. Consent: the customer must know what data is used, for what, and agree to it, under the personal data protection laws now in force across the region. Relevance: data must predict credit risk, not proxy for protected characteristics, which is where models trained carelessly on social or device data get into trouble.
Model design: accurate and explainable
Gradient-boosted models on well-engineered transaction features typically outperform both traditional scorecards and deep learning on tabular credit data, and they can be explained feature by feature with standard attribution methods. That explainability is not optional: regulators expect a lender to state the main reasons for a decline, customers have a right to understand decisions under the data protection laws, and the credit team needs to see why the model behaves as it does before it trusts it.
Design for the whole lifecycle. Train on the lender’s own outcomes where they exist and on carefully chosen proxies where they do not. Hold out by time, not at random, so that the validation reflects how the model will actually be used. Monitor stability and performance by segment after launch, because the population that applies changes when the acceptance policy changes. And keep a challenger model running so that improvements are measured rather than assumed.
Fairness and the regulator
Every regulator in the region has moved in the same direction: model risk management, documentation, human oversight and consumer protection apply to AI credit models as they do to any scorecard. Test for disparate outcomes across nationality, gender and age even where those variables are excluded, because other features can carry them. Document the data lineage, the training population, the validation and the limitations in a form a supervisor can read. Give customers a route to a human review. Lenders that build this in from the first model find it accelerates approval of the second and third.
From score to product
A better score only matters if the product can use it. Decisioning in real time at the point of need, inside the mobile app, at the merchant or in the SME platform; limits and pricing that respond to the score and to the customer’s behaviour over time; and collections that use the same data to distinguish a customer who cannot pay from one who forgot. The lenders winning the thin-file segment have built the scoring, the product and the journey as one system.
How Eurisko helps
Eurisko builds credit decisioning systems for banks and fintechs in the Gulf and Lebanon: data pipelines over transaction, bureau and alternative sources, explainable models with monitoring and challenger frameworks, and the lending journeys that use them, deployed inside the lender’s data boundary. If thin-file customers are a segment you decline today, our AI team can show you what your own data would support.
