AI-powered indicators quantify economic uncertainty from news
Researchers at the Banco de España have developed Retrieval-Augmented Uncertainty Indicators (RAUI), a new methodology using AI and newspaper data to generate topic-specific economic uncertainty measures. The approach leverages semantic search and large language models to quantify uncertainty, with applications for the Spanish economy.
AI refines uncertainty measurement
The RAUI methodology, inspired by Retrieval-Augmented Generation (RAG) systems, offers a novel approach to creating topic-specific uncertainty indicators from newspaper data.
It employs semantic search with an embedding model to identify news articles relevant to a user-defined topic, overcoming the limitations of traditional dictionary-based methods that struggle with context-dependent meanings and evolving language.
Once relevant articles are selected from large corpora, a large language model (LLM) quantifies the level of uncertainty within each.
This two-stage process allows for the nuanced interpretation of complete texts, distinguishing subtle linguistic variations that simple word-counting cannot.
The system maintains computational feasibility by only passing a subset of documents to the LLM, making it a significant advance for text-based sentiment and uncertainty analysis.
Uncertainty's varied impact on Spain
The RAUI indicators offer practical applications for the Spanish economy, demonstrated through two empirical analyses.
A vector autoregressive (VAR) model shows that internal sources of uncertainty create larger and more persistent effects on Spanish output than external ones.
Additionally, the indicators are used to construct time-varying fan charts around the Banco de España's GDP growth projections.
This reveals that forecast errors are systematically related to the level of uncertainty, with external uncertainty widening forecast error bands more significantly than internal uncertainty.
These applications underscore the indicators' ability to provide granular insights into economic dynamics.
Beyond keyword counting
This study represents a significant methodological leap, moving beyond the limitations of traditional dictionary-based approaches to capture nuanced economic sentiment.
While computationally intensive, its ability to generate topic-specific, context-aware indicators offers supervisors and policymakers a more granular view of economic risks.
The practical applications for the Spanish economy underscore its immediate relevance for enhancing forecasting and policy analysis.