The Bank of Italy published a study finding that credit sentiment indicators generated by large language models improve the ability of its In-house Credit Assessment System to distinguish solvent firms from insolvent ones. The research converts company related financial news from Dow Jones Factiva into quantitative indicators and tests their integration into the system, which forms part of the Eurosystem Credit Assessment Framework. The analysis uses models including BERT, LLaMA 3, Phi 3 and Gemma 2 to add structured insights from unstructured text to the Bank of Italy’s existing assessment approach. That approach combines a statistical model based on financial statements and National Credit Register data with expert judgment, which already considers qualitative online information but does not currently analyze it in a structured manner.
2026-09-16Bank of Italy
Bank of Italy study finds large language model sentiment indicators improve credit risk assessment
The Bank of Italy published research finding that large language model sentiment indicators derived from financial news improve its credit assessment system’s ability to distinguish solvent firms from insolvent ones. The approach adds structured analysis of qualitative information to the statistical and expert judgment components of the system.