Since 2014, FINMA has published anonymized summaries of every concluded enforcement proceeding. Collectively they form a decade-long record of regulatory intent: what FINMA treats as serious, what it considers structural, and how enforcement emphasis has shifted in response to the Credit Suisse crisis, the expansion of the crypto perimeter and a tightening geopolitical environment. Until now, that record has been readable only sequentially, one PDF at a time.

This project builds an end-to-end NLP pipeline that ingests years of German-language enforcement PDFs published by FINMA. Each case is encoded with a legal-domain sentence-transformer model, then fused with structured legal feature vectors: cited statutory provisions, sanction type and severity, entity category and conduct pattern. The result is a legal similarity embedding in which legally equivalent cases cluster together regardless of surface language variation, and a three-dimensional scatter in which a decade of enforcement action becomes spatially navigable.