Hodari Score explained: AI credit intelligence for African lenders
Reading mobile-money flow and repayment behaviour instead of formal credit history — and showing every factor.
A lender’s hardest decision is not whether to approve a good borrower or reject a bad one. It is what to do about the applicant they cannot see.
Across much of the continent, a large share of economically active adults have no meaningful record in a credit bureau. Not because they are bad risks — because their financial life has never touched the instruments a bureau records. They are paid in cash or by mobile money. They save in a group. They borrow from a supplier on terms nobody writes down. To a traditional scorecard, a decade of reliable trading looks identical to no history at all.
This is the thin-file problem, and it is not a data-collection problem. The data exists. It is simply not in the file the model was trained to read.
What alternative data actually means
Used carelessly, “alternative data” means scraping whatever a phone will give up and hoping the model finds something. That approach is both ineffective and, in our view, indefensible.
Used properly, it means a small number of signals that have a defensible causal story about repayment:
- Mobile-money flow — the rhythm, not the balance. Money moving in and out on a regular cadence describes a functioning business. Volatility that coincides with a known seasonal pattern is not the same as volatility that does not.
- Repayment behaviour on existing obligations — including informal ones the applicant chooses to evidence. Someone who has repaid twelve supplier advances has a track record, whether or not a bureau knows it.
- Tenure and stability — how long the same number, the same agent, the same counterparties have been in the picture.
- Declared context — what the loan is for, against what the business actually does.
What we deliberately do not use matters as much. Contact lists, message contents, browsing behaviour and social graphs can raise a model’s apparent accuracy while encoding exactly the kind of proxy discrimination that makes a lending book indefensible in front of a regulator. Accuracy bought that way is borrowed against a future problem.
Explainability is not a feature, it is the product
Any competent team can build a model that outputs a number. The number is the easy part. Three constituencies need to understand why, and none of them will accept “the model said so”:
The credit committee must be able to challenge a decision. A score they cannot interrogate is a score they will override, and a model that is routinely overridden is a model nobody is using.
The regulator will ask how the decision was reached, whether it can be reproduced, and what protects a protected characteristic from being inferred by proxy. “It is a neural network” is not an answer.
The borrower — who is a person being told no — is entitled to know what would have to change. A decline that comes with the two factors that drove it is a decline someone can act on. A decline with no reason is a customer lost permanently.
So every output carries its factor weighting: which signals moved the decision, in which direction, and by how much. Not a heat map — a sentence a loan officer can read aloud.
Where the hard problems are
We would rather be straight about these than pretend they are solved.
Proxy discrimination is subtle. A model never given gender or ethnicity can still learn a proxy for either from geography, device or transaction pattern. Testing for disparate outcomes across groups you did not model on has to be a standing process, not a launch checklist item.
Distribution shift is real. A model fitted on last year’s behaviour drifts when a currency moves, a fuel price moves, or a competitor changes their terms. Monitoring for drift and holding a retraining discipline is most of the operational cost.
Consent and data protection are load-bearing. You are processing sensitive financial data about identifiable people. The lawful basis, the retention period, the residency of the data and the applicant’s rights over it need to be settled before a single decision is made. In Ghana that means the Data Protection Act and the obligations that come with it, and it is not paperwork you retro-fit.
Some applicants should be declined. A scoring system whose purpose is to approve more people is a system that will eventually lend to people it should not, and the harm lands on the borrower first. The goal is to see more people accurately, which sometimes means seeing clearly that the answer is no.
What it looks like in practice
- Decisions returned fast enough to sit inside an application flow rather than beside it.
- Every decision reproducible: same inputs, same model version, same answer, with an audit trail.
- Policy controls the lender owns — thresholds, overrides and exclusions set by the credit team, not by us.
- Factor explanations attached to every decision, in language a customer-facing officer can use.
Where this stands
Hodari Score is in build. It is not available, and nothing here is a performance claim. The interface shown on our Technologies page is a design mockup and any figure in it is illustrative.
We are writing about it now because the design decisions above — what data to refuse, how much explanation to require, what to do about drift — are made at the start or not at all. If you are building or buying a scoring system, they are the questions worth arguing about, whoever builds it.