News-Based Inflation Expectations: LLM-Assisted Measurement and Forecasting

News-based measures of inflation expectations have been developed almost exclusively for high-resource languages such as English. Extending them to low-resource languages is limited by the high cost of expert annotation and the poor transfer of English-domain models. We develop a scalable, low-cost workflow that addresses these barriers—combining topic modeling, LLM-assisted labeling, and fine-tuned BERT classification—and apply it to construct a news-based inflation expectations index for Thailand. Drawing on 1.1 million Thai-language news articles from 2015 to 2024, the index leads both headline inflation and firms’ inflation expectations. It improves out-of-sample forecasting, reducing RMSE by up to 32% for headline inflation and 30% for firms’ expectations relative to an AR(1) benchmark, with gains increasing at longer horizons. Using SHAP-based decomposition, we uncover a horizon-dependent information structure: price-specific topics are most informative at short horizons, while broader macroeconomic narratives dominate at longer ones. We further show that inflation narratives are inherently subjective and often ambiguous, making downstream economic inference sensitive to prompt design. Our results demonstrate that LLM-assisted text analysis can deliver economically meaningful inflation indicators in low-resource, emerging-market settings, while underscoring the importance of careful prompt design.