Responsible scoring

AfriScore is a decision-support tool. It never lends, never holds the risk, and never makes the final decision. Approval, limit and pricing stay with the lender and its credit policy.

Every score comes with feature-level reasons and counterfactual suggestions that show what a borrower could change. Lenders and borrowers can both see why a score is what it is.

The engine returns a score, a risk tier and a probability of default. Your credit officers and committees decide what to do with them, and the pilot runs in shadow mode so nothing changes until you have evidence.

Fairness monitoring across region, device type and, where data permits, gender is planned as part of pilot calibration, with results to be published openly. It has not yet been measured, and we say so rather than claim otherwise.

The engine uses a defined set of alternative-data fields: mobile money summaries, utility payments, airtime patterns and device metadata. It can run inside the lender’s own environment with Docker, so borrower data does not have to leave.

We test feature ablation, data sparsity, distribution shift and the stability of explanations before any live lending use.

The engine was trained on synthetic data built from real mobile money distributions and has been piloted in Kenya. Markets such as the CEMAC region need real-data calibration first, and until then, published figures there are targets, not results.

If you believe a score or an explanation is wrong or unfair, or you want to raise a concern about how AfriScore is used, write to legal@afriscore.africa.