To assess an SME’s risk well, a scoring model needs to learn from many files. Yet no institution wants, or is allowed, to hand its clients’ data to a third party. Federated Learning solves this dilemma in banking and microfinance alike: the model travels, not the data.
The idea in one picture
Picture a teacher who never sees the students’ papers. They send an exercise to each class, each class practises on its own, then sends back only its general conclusions. By combining them, the teacher improves the lesson and sends it back to every class. No paper ever left its classroom.
With SCORE360, the classes are the financial institutions, the lesson is the scoring model, and the teacher is the ENGINE federated scoring engine.
One learning round, step by step
- Broadcast: SCORE360 sends the current version of the model to each participating institution.
- Local training: each institution trains the model in-house, on its own credit files.
- Encrypted sharing: only model parameters (weights and gradients, in other words numbers) are encrypted and sent to SCORE360. No names, no statements, no client files.
- Aggregation: SCORE360 combines these parameters into a more accurate global model, then sends it back.
What Federated Learning changes for your bank or MFI
First, you keep full control of your data: it is not copied, pooled or sold, as explained in our article on client data confidentiality. Second, you benefit from a model that has learned from the diversity of the whole network: other sectors, other SME profiles, other business cycles. Finally, the lending decision stays yours: the score is a decision-support tool, never a substitute for your credit committee.
Why it matters so much for African economies
In West Africa, SMEs drive most of the economy, yet access to credit remains a major challenge for them. Many are viable but hard to assess: few financial documents, little credit history, scattered data. Since the BCEAO rolled out PI-SPI, its interoperable instant payment platform, these businesses’ real activity leaves a digital trail. Federated Learning makes it possible to use this new wealth of information without ever centralising it.
A proven technology
Federated learning is not a lab concept. Google popularised it from 2016, notably to improve predictive text on smartphone keyboards without collecting what users type. It is now studied and used in sectors where confidentiality is critical, such as healthcare and finance. For the architecture in detail, see our federated learning technology.
In short
Federated Learning lets institutions that will never share their data build a more reliable score together. That is exactly what SME credit needs in the UEMOA region: more insight into risk, with no compromise on confidentiality, to finally finance the businesses that create jobs and growth.
FAQ
What is Federated Learning in banking?
It is a machine learning method in which the scoring model is trained inside each bank or MFI, on its own data. Only encrypted parameters are sent to a central engine, which combines them into a more accurate global model. Client data never leaves the institution at any point.
How is Federated Learning different from data sharing?
With traditional data sharing, client files are copied to a common database. With Federated Learning, the model travels instead: only weights and gradients, in other words numbers, are exchanged in encrypted form. No names, no statements and no client files are ever transmitted.
Is Federated Learning a proven technology?
Yes. Google popularised federated learning from 2016, notably to improve predictive text on smartphone keyboards without collecting what users type. It is now studied and used in sectors where confidentiality is critical, such as healthcare and finance, including for credit scoring.
Does a federated score replace the credit committee?
No. The score is a decision-support tool: it comes with its main explanatory factors so your analysts can justify their recommendation. The lending decision always stays with your institution and its credit committee, after human validation. The score informs the decision, it never makes it.