Research
AfriScore is a research-driven credit intelligence engine built to answer one question: how do you assess creditworthiness in markets where most adults have never interacted with a formal credit bureau?
In East, South, and Central Africa, a large majority of adults have no formal credit file, yet generate rich digital footprints through mobile money, airtime, and device usage. Traditional scoring models ignore this data because it doesn’t fit the "credit history" mould — leading lenders to either reject invisible applicants or rely on crude heuristics. AfriScore treats everyday digital behavior as a legitimate signal of creditworthiness.
Credit bureau coverage across much of the region remains low, particularly for informal-economy participants — even in markets with the most developed bureau infrastructure, a significant share of the adult population remains unscored.
Mobile money adoption across Sub-Saharan Africa is extensive, with hundreds of millions of registered accounts. That transaction data is a proxy for income stability, savings behavior, and financial discipline — the raw material AfriScore is built to use.
Because there is no large, labeled dataset of "good" and "bad" borrowers with mobile money footprints readily available, AfriScore uses a two-stage pipeline that separates pattern learning from probability estimation.
Stage 1
A synthetic population is constructed using empirical distributions derived from anonymized mobile money data patterns across East, South, and Central Africa. Key behavioral variables — inflow, outflow velocity, savings rate, airtime regularity, device type — are modeled to preserve realistic correlations, with a probabilistic default model assigning plausible repayment outcomes for training.
Stage 2
A Random Forest ensemble (200 trees, max depth 8, min samples split 50) captures non-linear interactions in the synthetic data. Isotonic regression (Pool Adjacent Violators algorithm) calibrates raw scores into well-calibrated probabilities of default as real repayment outcomes become available from pilot deployments.
Target
Target ROC-AUC
Target
Target Calibration Error
Target
Target Brier Score
These are engineering targets for the calibration phase in markets where AfriScore has not yet processed real repayment data, such as the CEMAC region. They will be replaced with measured results as CEMAC pilot calibration begins.
The model is designed to be evaluated against adversarial scenarios before any live lending use — missing data, noisy inputs, and distribution shift are all expected conditions in real deployment, not edge cases.
Measuring how much accuracy is lost when the most important individual features are removed.
Testing performance when a large share of input fields are missing or unavailable.
Testing whether the model generalizes across different regional behavioral distributions.
Confirming that explanation rankings remain consistent across resampled populations.
Every AfriScore decision is designed to ship with SHAP-based feature contributions and counterfactual explanations. Fairness monitoring across region, device type, and gender (where data permits) is planned as part of the pilot calibration process, with results to be published openly.
A book examining the structural foundations of financial systems across Africa — the broader thesis underlying AfriScore’s approach.
View on Amazon →The academic foundation for AfriScore’s scoring methodology and regulatory positioning.
Read on SSRN →Mapping the data, the market, and the policy frontier for alternative credit scoring across the region.
Read the policy paper →