A Sandbox-First Position on BCPNNs in EU Financial Services: Evaluate the Algorithm Now, Defer Production Deployment

Neuromorphic computing attracts considerable expectation in financial services and offers, so far, very little operational evidence. This position paper separates the two for one model family, Bayesian Confidence Propagation Neural Networks (BCPNNs), as implemented and tested in the Horizon Europe EXTRA-BRAIN project.

The position is that EU financial institutions should evaluate BCPNNs now as a hardware-agnostic algorithm, in controlled, non-production decision-support pilots, and should defer neuromorphic hardware procurement and any use in customer-affecting decisions or regulatory risk models until three evidence gates have been passed: finance-grade benchmarks, supervisor-usable explanations, and a model-governance protocol for continuously adapting estimators.

Three propositions are assessed separately, because the published evidence attaches to different layers: BCPNN as an algorithm on conventional hardware, FPGA acceleration of the same model, and a spiking BCPNN on neuromorphic silicon. The paper reviews the architectural rationale for BCPNN, the available tooling, the regulatory constraints under the EU AI Act, DORA and MiFID II, and the standardisation gaps in explainability, financial benchmarking and model-risk management. The EXTRA-BRAIN digital finance use case is presented as a sandbox and benchmark environment rather than as a production deployment, with falsification conditions stated in advance. Pre-normative input is offered to standards development organisations, and recommendations are given for financial institutions, supervisory authorities, standards bodies and funding authorities.