LuluFinance: Alternative Credit Risk Scoring via Mobile Deep Learning in Lagos

Nigeria's informal economy employs an estimated 80% of the country's workforce. Suya sellers, fabric traders, parts dealers, small-scale food processors — tens of millions of people participate in complex commercial networks, negotiate credit with suppliers, manage inventory, and build genuine businesses. Yet virtually none of them have a credit file in the conventional sense. They have no salary slips, no formal bank account history, no collateral that a bank recognises. The result is that the country's commercial banking sector serves a small minority of its actual economic participants, while the majority rely on rotating savings groups, informal moneylenders with usurious rates, or supplier credit extended at the cost of lower margins.

LuluFinance, incorporated in Lagos in 2020 by a former Goldman Sachs analyst and a machine learning engineer who previously worked at Flutterwave, set out to build a credit scoring system that does not require a formal financial history. Their approach, which has now processed over 1.2 million credit applications and extended over ₦18 billion ($11 million) in micro-loans, is built on a recursive neural network that ingests alternative data sources to construct a credit risk score without ever asking for a bank statement.

The Data Architecture

LuluFinance's scoring model ingests five primary data streams, all derived from the applicant's mobile phone with explicit consent: mobile money transaction ledger logs (amounts, frequencies, counterparty categories), SMS receipt patterns (identifying supplier payments, utility settlements, and customer receipts embedded in message text), airtime top-up frequency and amounts (a proxy for income regularity), app usage patterns on the loan applicant's device (cross-referenced against a database of apps associated with high-income versus low-income usage profiles), and geolocation data during business hours (verifying that the applicant regularly attends a declared business location).

The model does not use social media data, contact list analysis, or any data point that requires access to the applicant's communications content. This constraint was deliberate — not simply for regulatory compliance, but because early testing showed that social network proximity scores correlated strongly with existing social inequalities and produced a model that was accurate at aggregate level but systematically disadvantaged women, who on average have smaller and more geographically constrained social networks than men in the same income bracket.

The recursive neural network architecture was chosen over simpler gradient-boosted tree models because the sequential nature of transaction history — the rhythm of when money moves in and out, not just the volumes — contains predictive signal that architectures processing observations independently cannot capture. A trader who pays suppliers promptly on the first Friday of each month, receives customer payments clustered around the last week of the month, and maintains a consistent minimum airtime balance presents a different risk profile than a trader with similar average transaction volumes but chaotic timing — and the RNN captures that distinction in a way that feature-engineering for a gradient boosted model cannot easily replicate.

Performance and Fairness

LuluFinance's published model card, released in April 2024, reports a 90-day default rate of 8.2% across the full loan book — compared to an industry average of 23.5% for Nigerian digital lenders operating without alternative data scoring, according to the Digital Lenders Association of Nigeria's 2024 Annual Report. Loan sizes range from ₦20,000 to ₦500,000 ($12 to $300), with a median of ₦85,000 ($52), calibrated to the working capital cycle of typical informal market traders.

The fairness audit, conducted independently by a Lagos-based AI ethics consultancy, found that the model's false positive rate (incorrectly classifying a creditworthy applicant as high-risk) differed by less than 2% between male and female applicants and by less than 3.5% between applicants in the six geopolitical zones. The consultancy noted, however, that the model has not been tested on applicants from internally displaced populations or nomadic communities, for whom the geolocation verification component would produce unreliable results, and recommended that LuluFinance either exclude these populations from algorithmic scoring or develop a separate model validated on their specific data patterns.

The Regulatory Context

The Central Bank of Nigeria's Digital Finance Policy, issued in 2022, requires digital lenders to disclose the general nature of their credit scoring methodology and to provide applicants with a reason for loan rejection. LuluFinance complies with both requirements: rejected applicants receive an SMS message explaining, in plain language, the primary factors that led to the decision, and the company's website hosts a summary of its alternative data scoring approach.

What the policy does not require — and what consumer advocates have called for — is the right to contest a credit decision by providing additional evidence, the ability to request a human review of an algorithmic rejection, or independent auditing of the model's performance across demographic groups. LuluFinance has committed to publishing an annual model card voluntarily, but the company is one of over 400 licensed digital lenders in Nigeria, and most of those 400 publish nothing about their scoring methodology at all.

The case for AI-driven alternative credit scoring in Nigeria's informal economy is compelling. The case for ensuring that the models driving those decisions are transparent, fair, and subject to meaningful oversight is equally compelling. Right now, the first case is winning the argument considerably more easily than the second.