Bank for International Settlements publishes research using large language models to build and decompose US growth and inflation sentiment from news

The Bank for International Settlements published a working paper proposing an approach that uses large language models (LLMs) to parse press narratives directly and construct daily sentiment indices for US growth and inflation, with an explicit decomposition into underlying demand and supply drivers. The paper argues the indices provide a higher-frequency, near real-time read on macro conditions, track conventional “hard data” closely, and can improve the forecasting performance of simple statistical models. The methodology applies prompting-based classification to roughly 200,000 Wall Street Journal articles from January 2000 to April 2025 (sourced via Factiva), screening and then classifying relevant items using GPT-4.1 mini and GPT-5 mini. Articles are scored for directional sentiment and mapped into a taxonomy that splits sentiment into demand versus supply drivers, then into subcomponents such as real versus financial demand (including fiscal and monetary policy subdrivers) and supply factors including commodity prices, supply disruptions and government policy. A manual review of 300 classified articles yielded an estimated 92% accuracy, with directional sign errors reported as rare. The resulting indices are presented as consistent with business-cycle and inflation episodes and show leading properties in correlations with activity and inflation measures, while demand and supply decompositions are compared with model-based benchmarks cited in the paper. In out-of-sample tests, adding sentiment indices to AR(1) and local projection forecasts generally reduces RMSE at short to medium horizons across multiple growth and inflation measures, with gains varying by series and horizon. The paper states that the resulting sentiment indices will be made available on the BIS website.