News sentiment indicator significantly improves GDP growth forecasts
Swiss National Bank researchers developed a resource-efficient methodology to measure economic outlook in news, significantly improving GDP growth forecast accuracy. The new indicator, NEOS, captures sentiment shifts weeks before official releases.
Combining LLMs with document embeddings
Researchers at the Swiss National Bank (SNB) have developed a novel, resource-efficient methodology for measuring economic outlook in news text.
This approach strategically combines high-dimensional document embeddings from a specialized transformer model with synthetic training data generated by large language models (LLMs).
A logistic regression model, trained on these synthetic articles, is then applied to real news article embeddings to produce economic outlook scores.
This multi-step process enables local sentiment classification, addressing critical institutional constraints such as data usage agreements (DUAs) that restrict transmitting proprietary news content to external services.
The methodology requires minimal computational resources compared to direct LLM classification, making it compatible with internal infrastructure limitations and ensuring data confidentiality.
Outperforming traditional forecasting tools
The resulting News-based Economic Outlook for Switzerland (NEOS) indicator significantly improves GDP growth forecast accuracy, outperforming both survey-based benchmarks and traditional dictionary methods.
Timelier versions of NEOS, computed from partial monthly data, capture sentiment shifts weeks before official releases, proving particularly valuable during crises.
The indicator's interpretability allows for the identification of specific drivers of economic sentiment and enables sector-specific analyses, offering deeper insights into economic outlook before official data becomes available.