Every institution knows its clients very well. But it only knows them. An MFI focused on retail trade has few examples in agribusiness; a bank oriented towards large companies sees few very small businesses. Its credit scoring model inherits these blind spots.
The blind spots of a credit scoring model
A model learns what it is shown. If it has never observed a sector, a season or a business profile, it assesses it poorly: either it turns down good files out of caution, or it accepts risks it cannot read. Either way the institution loses: loans not granted, or defaults.
What the network brings
In a federated learning network, each institution contributes what it knows best. The global model combines these experiences: it recognises risk patterns observed elsewhere, in other sectors or other UEMOA countries, then sends that knowledge back to everyone.
A network across the UEMOA
PI-SPI connects banks, electronic money issuers and microfinance institutions across the eight UEMOA countries around a single instant payment system. SME flows therefore follow common formats from one country to the next. It is an ideal foundation for regional federated learning: an MFI in Dakar and a bank in Abidjan can contribute to the same model, each keeping its data at home.
A benefit that costs no data
Network effects are usually reserved for players who centralise data. Here, they come without pooling: each institution keeps its files and shares only encrypted parameters, never its client data. Joining the network is not giving up a competitive edge, it is gaining one. And the more credit flows to SMEs, the more the regional economy benefits.
A model that stays current
SME activity changes: farming seasons, price swings, new mobile payment habits. Because learning repeats round after round, the model absorbs these changes continuously instead of freezing at the moment it was designed.
What about your specific needs?
Benefiting from the network does not mean giving up your risk policy. The score computed by the ENGINE federated scoring engine comes with its main explanatory factors, and your institution sets its own thresholds and rules, and takes the final decision.
FAQ
How can a credit scoring model become more accurate?
A model learns what it is shown. To become more accurate, it needs to observe more sectors, seasons and business profiles. Federated learning makes this possible by combining the experience of several banks and MFIs, without any of them sharing client files, and by retraining the model round after round.
What is the network effect in Federated Learning?
Each institution trains the model on its own data and contributes to a global model. The more institutions the network has, the more situations the model has seen, and the sharper it becomes for all. This effect comes without pooling: only encrypted parameters are shared between participants.
Does joining a federated network mean giving up a competitive edge?
No. Each institution keeps its files, its clients and its risk policy. It shares only encrypted parameters, which cannot be used to rebuild a file. In return, it benefits from a model that has learned from other sectors, other SME profiles and other UEMOA countries.
Does the global model respect my own risk policy?
Yes. The score comes with its main explanatory factors, and your institution sets its own thresholds and rules and takes the final decision. The global model informs your teams’ analysis without ever replacing your credit committee or your lending policy.