The 190 Million Who Are Creditworthy But Invisible
EU formal credit bureau infrastructure covers approximately 450 million individuals — a remarkable achievement in a generation, but still leaving close to 190 million adults outside the system's view. Among those 190 million are some of the most creditworthy borrowers in the country: first-generation entrepreneurs who have never needed formal credit, women who manage household and business finances but are not the primary account holder, SME operators who run lean cash businesses and have no credit history because they have never defaulted because they have never borrowed.
The conventional response to thin-file applicants is rejection — not because these borrowers are risky, but because the risk measurement infrastructure cannot see them. This is not a risk management decision. It is a data availability decision masquerading as a risk management decision. The Thin-File AI replaces the missing bureau data with something more immediately informative: the actual financial behaviour of the borrower, observable in real time through VAT filings, SEPA Instant transaction records, and banking patterns that together tell a more complete credit story than a bureau file built on historical borrowing alone.
The Signal Architecture: 4 Data Categories, 28 Computed Metrics
Business Revenue & Compliance Signals — Highest Weight for SE Borrowers
Weight: 32%Real-Time Cash Flow & Payment Behaviour — Highest Frequency, Current Signal
Weight: 28%Savings, Liquidity, and Financial Discipline Signals
Weight: 26%Supporting Signals — Corroboration and Context
Weight: 14%The Credit Score Output: What the AI Produces
How the Thin-File Score Is Calibrated Against Actual Default Data
A credit score that is not validated against actual default outcomes is not a credit score — it is a heuristic. The Thin-File AI is trained and calibrated on a dataset of borrowers who were initially thin-file at origination, whose loans have matured sufficiently to generate 90-DPD outcome data, and whose alternative data signals at origination can be retrospectively compared to their repayment behaviour. This outcome-linked calibration is what allows the Thin-File Score to be mapped to a bureau-equivalent risk band rather than existing as an opaque proprietary metric that lenders cannot benchmark.
The calibration shows that thin-file borrowers with a TFS above 700 have a 12-month default rate of 2.8% — broadly comparable to bureau-scored borrowers with a credit score of 700–730. The correlation between VAT turnover consistency and subsequent repayment is particularly strong in the self-employed segment: borrowers with 8+ quarters of consistent or growing VAT turnover default at 1.9%, compared to 4.8% for those with volatile or declining turnover — a difference that bureau-only underwriting cannot detect because it has no VAT data.
Alternative Data Is Not a Concession to Risk — It Is a Correction to a Measurement Gap
The instinct to treat thin-file lending as inherently riskier than bureau-scored lending is understandable but factually incorrect. Thin-file borrowers are not more likely to default — they are less measurable by traditional instruments. The Thin-File AI solves the measurement problem, not the risk problem. A borrower who has demonstrated six years of consistent business cash flow, timely tax compliance, and savings discipline is not a high-risk borrower. She is a low-risk borrower whom the traditional system cannot see. The Thin-File AI makes her visible — and makes the case for her credit, with evidence.