The thin-file problem is real and the response to it has been reasonable: if a founder has no bank history, build a score from what can be observed, and lend against that. Across the continent this has produced a generation of alternative scores drawn from airtime top-ups, mobile-money velocity, phone contact graphs, and, inside accelerators, from programme behaviour itself. The reasoning is sound and the execution keeps drifting in the same direction, because the easiest things to observe inside a programme are attendance, punctuality, assignment completion and staff impressions, and a score assembled from those measures how well a founder does programme, not how likely they are to repay. Those are different variables. Conflating them produces a number that feels rigorous, correlates with conscientiousness, and tells a lender remarkably little about the thing they are actually underwriting.
Key Takeaways
- Programme behaviour and repayment probability are different variables. A score built from attendance measures compliance, not creditworthiness.
- The strongest available signals are transactional: money that already moved, on a rail that can be checked, over a period long enough to show a pattern.
- Any score used for lending must be predictive against observed outcomes, and most founder scores have never been back-tested against a single default.
- Scoring a founder for a place in a programme and scoring them for a loan are two instruments. Building one and using it for both is the common design error.
- A score that cannot be explained to the founder it declined is not a score, it is a decision with decoration.
- The first thing to measure is not the founder’s worthiness. It is whether the instrument itself fits: debt is the wrong shape for pre-revenue hardware regardless of the score.
What signals actually carry weight?
Money that moved, verified on a rail somebody else keeps. Mobile-money inflows over twelve months carry more information than any self-report, because they are dated, third-party recorded and difficult to dress up. Supplier payment history carries similar weight where it can be obtained. Signed offtake agreements and paid invoices sit above pipeline and letters of intent, and far above projections, because they involve a counterparty who has already accepted a cost. The credibility hierarchy is the underwriting question more than it is the pitching question.
Below that tier, behavioural signals have genuine but narrower use. Repayment of a small prior facility is the best predictor of repayment of a larger one, which argues for graduated lending rather than better prediction. Consistency of small deposits indicates cash discipline. Beyond those, the marginal signal drops quickly, and the temptation to keep adding variables should be resisted, because each additional weak variable makes the score feel more sophisticated and perform slightly worse.
What does not belong: attendance, assignment quality, staff warmth, and articulacy. All four correlate with education and confidence, both of which correlate with class, and none of which are what is being underwritten. A score that quietly proxies for social background will look predictive in a small sample because the same founders do well on everything, and will fail exactly when the programme succeeds at widening access.
Why do programme scores and lending scores need to be different instruments?
Because they answer different questions and carry different costs when wrong. A selection score asks: will this founder finish the programme and contribute to the room? Getting that wrong costs a seat. A credit score asks: will this founder repay under stress? Getting that wrong costs the loan, the relationship, and often the founder’s business.
The variables genuinely differ. Business status at entry predicts completion strongly, because an existing venture supplies the reason and rhythm that carry someone through a four-month evening commitment. It predicts repayment far more weakly, because repayment turns on cash conversion and shock absorption rather than motivation. Community contribution is a legitimate selection criterion and close to irrelevant to underwriting. Ability to pay a course fee is not ability to service a facility.
The design error is building one instrument, validating it loosely against programme outcomes, and then pointing it at a loan book because it is the number the institution already has. The cheaper correction is to keep two scores with an explicit divider, and to state on the face of each which decision it is licensed for.
How do you know the score works?
You back-test it, and almost nobody does. A score is a prediction, and an unvalidated prediction is an opinion in a numeric costume. The minimum honest standard is to hold the historical file, run the score retrospectively against founders whose outcomes are already known, and check whether the score separated them. If the defaulters and the repayers score alike, the instrument does not work, however carefully it was designed.
This is uncomfortable in a small portfolio, because the numbers are thin and the failures are people you know. It is also where the discipline earns its keep. A programme with seven loans disbursed and a meaningful share at risk has enough information to learn something specific, and refusing to look is a choice. The scoring model is not the asset. The back-test against actual outcomes is the asset, and until it exists the responsible use of the score is advisory rather than decisive.
FAQ
Is alternative credit scoring a bad idea?
No. It is the right response to thin files. The failure is in variable selection and validation, not in the concept.
What if we have too few loans to back-test?
Then use the score to inform a human decision rather than to make it, disclose that it is unvalidated, and lend in graduated steps so the portfolio generates the evidence.
Should founders see their score?
Yes, with its components. A score you cannot explain to the person it declined will not survive its first serious challenge, and explaining it usually improves it.
Does a good score justify a loan?
Not on its own. The prior question is whether debt fits the venture at all. A pre-revenue manufacturer with a strong score is still the wrong candidate for monthly-interest debt.
Related Reading
- Who actually finishes: what 99 alumni say about selection
- Founders do not lack data, they lack a decision the data could change
- Patient capital for manufacturing founders
- The missing middle is a management problem, not a capital problem
Sources and Evidence
- Blacksmith, N. and McCusker, M. E., Data-Driven Decision Making in Entrepreneurship, on validation, variable selection and the limits of behavioural proxies.
- Internal alumni outcomes dataset, 99 participants 2022 to 2026, including seven disbursed facilities, showing completion and lending outcomes as distinct variables.
- Baig, P., Montoya-Martinez, E. et al., Entrepreneurial Ecosystems: Resource Acquisition, Flow and Management, on resource allocation and the mismatch between instrument and venture stage.