Assessing a borrower with no banking history
Refusing for lack of information is not a risk decision: it is the absence of one. Three principles make a thin file assessable without inventing what is not known.
Refusing by default has a cost — it is simply invisible
An institution measures precisely what it approved wrongly: non-performing loans, provisions and recovery costs all appear in its accounts. It never measures what it refused wrongly, because that loss appears nowhere.
This asymmetry of measurement produces an asymmetry of behaviour. Unable to assess a trader whose flows run through mobile money, or a business whose bookkeeping is rudimentary, the institution refuses — and never learns what the refusal cost.
The financing gap for African small and medium businesses is estimated at 331 billion dollars by the International Finance Corporation. Part of that gap reflects not the real risk of borrowers, but the impossibility of measuring it.
First principle: absence is information
The most widespread practice when a value is missing is to substitute an average. That is convenient for a statistical model, and wrong for a credit decision.
A file with no proof of income is not a file with average income: it is a file whose income is unknown. Those two situations call for opposite treatments — the first leads to a score, the second to a document request.
Treating absence as absence has an immediate practical consequence: the system can state a confidence level, and route a file that is too incomplete to human review rather than decide it automatically.
Second principle: start from a scorecard, not from a model
A statistical model needs thousands of files whose outcome is known. An institution that has never lent to a segment has, by definition, no history on that segment. Waiting to have one means never starting.
An expert scorecard sidesteps the deadlock: it states the credit policy the institution already applies, with its thresholds and weights, in an explicit and testable form. It does not claim to be optimal; it claims to be faithful and verifiable — which is enough to begin.
The learned model takes over as history accumulates. The share of each becomes a setting the risk department moves knowingly. The scorecard remains as a guardrail and as a reference for explanation.
Third principle: recognise the collateral actually used
Systems designed for other markets almost always reason on real-estate collateral. Yet credit here is secured through other mechanisms whose predictive value is established in practice: joint liability guarantees, progressive lending cycles where the amount rises with repayment history, inventory pledges.
A model that cannot represent these mechanisms does not merely ignore them: it structurally penalises the files that rely on them — which are precisely the files one is trying to assess.
The observable regularity of flows — credit movements, punctuality of recurring payments, length of trading history — is real assessment material. It already exists in the institution's systems, and it is rarely used.
Sources
- International Finance Corporation — estimate of the African SME financing gap.
- World Bank — Global Findex, unbanked adult population in sub-Saharan Africa.