Generic, US-trained fraud models perform poorly on African transaction data — the signal that predicts fraud in one market (say, a first-time card used from a new device) is completely ordinary in a market where most cards are shared within a household.
Our fraud models are trained separately per market on the transaction behavior actually observed there, and re-trained on a rolling basis as patterns shift. Every charge is scored in under 40ms, before the payment is authorized, using a combination of device signals, velocity checks, and channel-specific heuristics.
Businesses can tune their own risk thresholds in the Mereb Hub — stricter thresholds catch more fraud at the cost of some false declines; looser thresholds optimize for conversion. We publish suggested defaults per industry as a starting point.