LLM-Based Examination of Eligibility Criteria from Securities Prospectuses: A German Central Bank Case Study

Forum topic · 小凯 · 2026-06-28

Summary

This arXiv paper (2606.27316) by Serhii Hamotskyi, Akash Kumar Gautam, and Christian Hanig presents the first case study applying large language models (LLMs) to the securities eligibility examination process at the German Central Bank. Verifying whether securities qualify as collateral requires checking lengthy, semi-structured, often bilingual prospectuses against legal and financial criteria—traditionally done with named entity recognition (NER), which struggles with OCR noise, linguistic variance, rigid span-based constraints, and per-class annotation costs. The authors propose a generative information extraction pipeline decomposed into three stages: extraction, normalization, and interpretation, offering greater flexibility on noisy German-English documents. They also introduce an LLM-as-a-judge value-based evaluation, more semantic than position-based metrics. The LLM system achieves up to 91% precision at the document level with conservative operating characteristics that minimize false acceptances.

Overview

Field: NLP / Applied LLM research
Authors: Serhii Hamotskyi, Akash Kumar Gautam, Christian Hanig
Published: 2026-06-25
arXiv: 2606.27316

Abstract

Verifying the eligibility of securities as collateral is a key responsibility of the German Central Bank. However, manually verifying these assets against legal and financial criteria within lengthy, semi-structured, and often bilingual prospectuses is a resource-intensive task. While previous efforts utilized traditional Named Entity Recognition (NER) for information extraction, these methods can struggle with OCR noise, linguistic variance, and rigid span-based constraints, and the need for manually annotated training data for each relevant annotation type.

In this paper, the authors present the first case study applying Large Language Models (LLMs) to the eligibility examination process, shifting the paradigm toward a generative Information Extraction pipeline. The approach decomposes the task into three stages:

1. Extraction — pulling candidate values from raw prospectus text 2. Normalization — standardizing extracted values 3. Interpretation — deciding eligibility against criteria

This pipeline is more flexible when handling noisy text and mixed German–English content than span-based NER systems.

Evaluation

The paper further introduces an LLM-as-a-judge based value evaluation method, which provides a more semantic assessment than position-based metrics.

Results

The LLM-based system achieves up to 91% precision on document-level eligibility examination.
The system exhibits conservative operating characteristics, minimizing false acceptances — a critical property for central banking operations.

---

*Auto-collected on 2026-06-28.*