Measuring economic sentiment from open-ended survey comments using large language models

Abstract

This article develops a novel economic sentiment indicator (LLM-ESI) by applying large language models to open-ended responses from Swiss business tendency surveys. Using a BERT-based transformer model, it extracts firm-level sentiment from free-text survey comments and aggregates it into a high-frequency indicator of macroeconomic conditions. The LLM-ESI closely tracks the business cycle and performs on par with, or better than, traditional benchmarks in nowcasting GDP. These results highlight the potential of large language models and open-ended survey responses to deliver timely and nuanced signals for real-time economic analysis.