Federated Modelling: A new framework and an application to system-wide stress testing

This paper builds on existing literature on federated learning to introduce an innovative framework, which we call federated modelling. Federated modelling enables collaborative modelling by a group of participants while bypassing the need for disclosing participants’ underlying private data, which are restricted due to legal or institutional requirements. While the uses of this framework can be numerous, the paper presents a proof of concept for a system-wide, granular financial stress test that enables effective cooperation among central banks without the need to disclose the underlying private data and models of the participating central banks or their reporting entities (banks and insurers). Our findings confirm that by leveraging machine learning techniques and using readily available computational tools, the framework allows participants to contribute to the development of shared models whose results are comparable to those using full granular data centralization. This has profound implications for regulatory cooperation and financial stability monitoring across jurisdictions.