LLM-Based Examination of Eligibility Criteria from Securities Prospectuses at the German Central Bank

Research Approach

This paper introduces a generative Information Extraction (IE) pipeline designed to automate the verification of securities eligibility for the German Central Bank. The process replaces traditional, span-based Named Entity Recognition (NER) models with a multi-stage LLM architecture that performs extraction, normalization, and interpretation. By utilizing high-performance models like Llama-3.3-70B-Instruct and Cohere Command-R, the system handles semi-structured, bilingual (German-English) prospectuses that are often plagued by OCR noise and complex formatting.

Methodology

To overcome the limitations of rigid, location-based evaluation metrics, the authors implement a value-based evaluation methodology using 'LLM-as-a-judge.' This approach assesses the semantic correctness of extracted information rather than relying on exact text matches or character offsets. The pipeline processes documents in three distinct stages: extracting raw data, normalizing it into a structured format, and interpreting that data against specific eligibility criteria (e.g., currency, instrument type, and subordination status). The system also incorporates master data and base prospectus annotations to ensure robust decision-making.

The LLM-based system significantly improves upon previous NER-based approaches, particularly for linguistically complex criteria such as redemption terms and debt subordination. The results show high precision (up to 91%), with the system exhibiting a conservative operating profile that is well-suited for financial risk management, as it prioritizes avoiding false acceptance of ineligible assets. The study also highlights that smaller, specialized models like Command-R (32B) can perform competitively with much larger models, suggesting that architectural efficiency and grounded generation capabilities are critical for financial NLP tasks.

Manual verification of financial prospectuses is a resource-intensive bottleneck for central banks. This research provides a scalable, flexible framework that reduces the need for extensive manual annotation while improving accuracy in noisy, real-world document environments. By shifting toward generative pipelines and semantic evaluation, the authors provide a blueprint for automating complex compliance tasks in the financial sector.