LLM news sentiment improves central bank credit risk assessment
Banca d'Italia researchers Manuel Cugliari and Giulio Gariano found that integrating large language model sentiment indicators into in-house credit systems sharpens corporate default detection. The September 2026 study applies models like LLaMA 3 and BERT to Dow Jones Factiva news.
From news flow to credit scores
The research team utilized four state-of-the-art language models—BERT, Meta's LLaMA 3, Microsoft's Phi 3, and Google's Gemma 2—to extract credit sentiment indicators from financial articles in the Dow Jones Factiva database.
These indicators were embedded directly into Banca d'Italia's In-house Credit Assessment System (ICAS), a framework operating within the Eurosystem Credit Assessment Framework (ECAF) to evaluate bank loans pledged as monetary policy collateral.
The empirical results demonstrate that sentiment scores derived from unstructured textual data systematically improve the ability to distinguish solvent firms from insolvent ones relative to purely statistical balance sheet metrics.
Structuring qualitative analyst judgment
Under Eurosystem rules, central banks assess collateral eligibility using rating agencies, internal bank models, or domestic ICAS platforms.
Banca d'Italia's ICAS combines a statistical model estimating one-year default probabilities with expert qualitative assessments.
While analysts regularly review press reports, qualitative data previously lacked quantitative structure.
Automated sentiment scoring bridges this gap by systematically standardizing narrative corporate information.
Promising automation, persistent black box
Automating narrative analysis reduces analyst bias across large loan portfolios.
Yet central banks cannot treat LLM sentiment as objective truth without guarding against model hallucinations.
Real-world collateral frameworks will require strict interpretability over pure statistical gains.