LLMs detect 79 capital rule divergences in G-SIB AT1 prospectuses
A Bank for International Settlements study presents a large language model screening procedure to identify divergences between bank prospectuses and regulatory capital rules. Testing across ten European global systemically important banks identified 79 recurrent candidate discrepancies.
Seventy-nine flags across ten banks
The procedure compares regulatory texts with issuance prospectuses using structured prompts across five independent runs to account for model variability.
In a review of Additional Tier 1 instruments from ten European global systemically important banks across the European Union, the United Kingdom and Switzerland, the system flagged 79 recurrent candidate divergences appearing in at least four runs.
Provisions governing mechanical triggers and point of non-viability accounted for roughly two-thirds of the flagged differences.
The identified variations fall into five specific categories: narrowing, expansion, omission, ambiguity and contradiction.
These flags prioritize clauses that determine when capital loss absorption begins.
Validation through historical stress
Validation on anonymized historical documents successfully recovered contractual divergences that later triggered major legal disputes.
The model flagged the narrowed public support condition in Credit Suisse prospectuses prior to the CHF 16.5 billion write-down in 2023.
It also detected the loss-absorption sequencing contradiction in Yes Bank's 2013 bonds before the 2020 restructuring.
The methodology offers applications for assessing domestic rules against Basel standards under international review programs.
Promising triage, unproven judgment
Automated document screening solves a real resource bottleneck by pinpointing subtle contractual deviations before stress hits.
Yet recurrence across model runs cannot replace legal reasoning on nuanced edge cases.
Supervisors must establish rigorous error benchmarks before deploying these tools in binding regulatory reviews.